diff --git a/Cargo.lock b/Cargo.lock
index d4df3417..3cb4580d 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -4884,6 +4884,7 @@ dependencies = [
  "hive-jobq-metrics",
  "hive-jobq-wire",
  "hive-log",
+ "hive-sh4re",
  "hive-types",
  "hmac 0.13.0",
  "http",
@@ -5014,6 +5015,7 @@ dependencies = [
  "axum",
  "data-encoding",
  "futures-util",
+ "hive-sh4re",
  "reqwest",
  "rustls",
  "serde",
diff --git a/docs/agent-lifecycle/config-changes.md b/docs/agent-lifecycle/config-changes.md
index 0cf83813..59b2a9c2 100644
--- a/docs/agent-lifecycle/config-changes.md
+++ b/docs/agent-lifecycle/config-changes.md
@@ -253,8 +253,9 @@ the operator review and merge it. By design, no second, mount-shaped
 path reaches the same file without the review.
 
 Agents holding the `manage_root_agent` capability (granted per agent in
-`capabilities.json`; see `hive-sh4re/src/permissions.rs`) get additional
-host-side bind mounts via `set_nspawn_flags`:
+the swarm's `agent-permissions` record; see
+`docs/process/conventions.md#capabilities`) get additional host-side bind
+mounts via `set_nspawn_flags`:
 
 - `/var/lib/hyperhive/agents/` → `/agents/` (RW) — **every** agent's
   proposed repo, not just direct children. The capability means "may
@@ -267,16 +268,13 @@ host-side bind mounts via `set_nspawn_flags`:
 An agent without the capability only has its direct children's config
 dirs.
 
-⚠️ nspawn bakes bind flags at container start, so granting or
-revoking this capability doesn't change any mount until that agent's
-container is rebuilt/restarted.
+⚠️ nspawn bakes bind flags at container start, so hive-c0re restarts
+an agent, after its current turn, when its capabilities change.
 
-The root agent gets the capability by default, seeded on its autodeploy
-path (`workers::auto_update::ensure_root_agent`) so the recovery mounts
-are there from its first container. That seed only fires while
-`capabilities.json` doesn't exist yet: any grant or revoke through
-the dashboard creates the file, so a revoked root-agent grant stays
-revoked and isn't re-applied on the next hive-c0re restart.
+The root agent holds the capability by default: an agent whose record
+leaves `capabilities` unset gets the hive's default, which is
+`manage_root_agent` for `ruth` and nothing for every other agent.
+Setting `"capabilities": []` for her revokes it.
 
 Each proposed repo (`/agents/<n>/config/`) is pre-configured
 with `applied` as a git remote pointing at
diff --git a/docs/agent-lifecycle/persistence.md b/docs/agent-lifecycle/persistence.md
index 1dbc2cb8..ce507539 100644
--- a/docs/agent-lifecycle/persistence.md
+++ b/docs/agent-lifecycle/persistence.md
@@ -391,10 +391,6 @@ Contents:
   (`{ "alice": ["messaging", "inbox", "execution"] }`). Written by
   `tool_groups::set_groups`; injected as `HIVE_TOOL_GROUPS` env
   var into each agent's container.
-- `capabilities.json` — per-agent capability grants
-  (`{ "ruth": ["manage_root_agent"] }`). Written by
-  `capabilities::set_caps`; injected as `HIVE_CAPABILITIES` env
-  var. Absent agents have no extra capabilities.
 - `swarm-resource-limits.json` — the caps the swarm sets for this
   hive's agents
   (`{ "sock": { "cpu_quota": "400%", "memory_max": "8G" } }`).
@@ -418,6 +414,13 @@ Contents:
 
 The root agent has the meta dir RO-mounted at `/meta/`.
 
+Outside the meta repo, `/var/lib/hyperhive/swarm-agent-permissions.json`
+holds the capabilities the swarm grants this hive's agents, as
+`workers::swarm_permissions` last read them from the swarm's
+`agent-permissions` bucket (every 30 seconds): a copy, not the
+record, which `swarmctl agent permissions` sets. `set_nspawn_flags`
+reads it at every container start.
+
 hive-c0re renders containers onto `meta#<n>` at creation.
 
 ## Destroy vs purge
diff --git a/docs/process/conventions.md b/docs/process/conventions.md
index de91633f..31683462 100644
--- a/docs/process/conventions.md
+++ b/docs/process/conventions.md
@@ -334,35 +334,37 @@ that allows the underlying resource access.
 
 | Capability | Effect |
 |---|---|
-| `manage_root_agent` | may manage *any* agent: bind-mounts every agent's state (rw) + config (ro) into the holder's container, plus `/applied` and `/meta` (ro). Takes effect on the holder's next container rebuild/restart |
+| `manage_root_agent` | may manage *any* agent: bind-mounts every agent's state (rw) + config (ro) into the holder's container, plus `/applied` and `/meta` (ro). Takes effect when the holder's container restarts, which hive-c0re does itself on a change |
 
-**Config storage** — per-agent capabilities live in
-`/var/lib/hyperhive/meta/capabilities.json` alongside `tool-groups.json`.
-Format: `{ "ruth": ["manage_root_agent"] }`.
-An absent entry means "no extra capabilities." `render_flake` in `meta.rs`
-reads this file and injects `HIVE_CAPABILITIES` (comma-separated
-`snake_case` names) into each agent's systemd service env; absent entries emit
-no env var so agents without capabilities don't trigger a spurious rebuild.
+**Where they live** — in the swarm, not on the hive: the `capabilities` field
+of the agent's key in the `agent-permissions` KV bucket
+(`swarm-queue-client/src/agent_permissions.rs`), which swarm-controller alone
+writes. Set them with `swarmctl agent permissions get|set|clear <agent>`
+(`/api/agents/{name}/permissions`). Unset means the hive's default, which is a
+constant in hive-c0re (`capabilities::default_for`): none, except `ruth`, who
+gets `manage_root_agent`. `"capabilities": []` grants none.
 
-**Setting capabilities** — the operator sets capabilities via the
-C4P4B1L1T13S section in the dashboard's P3RM1SS10NS tab.
-`hive-c0re::capabilities::set_caps(name, caps)` is the write path.
-After a change `meta::sync_agents` commits the updated file; the next agent
-rebuild picks up the new `HIVE_CAPABILITIES` env var.
+**How a change applies** — every 30 seconds hive-c0re reads the key of each
+agent it runs (`workers/swarm_permissions.rs`) and stores the result in
+`/var/lib/hyperhive/swarm-agent-permissions.json`, outside the meta repo. nspawn
+fixes the bind mounts at container start, so when an agent's capabilities
+change hive-c0re restarts it after its current turn. A failed read keeps the
+last stored value; it never falls back to the default.
 
-**Runtime resolution** — at session start the harness reads `HIVE_CAPABILITIES`
-and resolves each token to a `Capability` variant, logging and skipping
-unrecognised ones. An absent or empty var means no extra capabilities.
+**Moving off `meta/capabilities.json`** — on its first boot with this code a
+hive copies the file's grants into the stored value, sends them to
+swarm-controller once (written only for agents with no key, and only for
+agents that hive's wanted state declares), and deletes the file from the meta
+repo when the controller answers (`workers/permissions_import.rs`).
 
 **Capability NOT configurable from `agent.nix`** — same reasoning as tool
 groups: an agent that could grant its own capabilities via a config commit would
 bypass the operator approval gate.
 
 **Adding a new capability** — add a variant to `Capability` in
-`hive-sh4re/src/permissions.rs` + an arm to `as_str`. Add it to `Capability::ALL`
-(the source of truth for the permissions UI columns). Implement the access
-check in the relevant handler (`hive-c0re/src/socket_server/mod.rs`,
-`coordinator.rs`, or a handler under `hive-c0re/src/dashboard/`).
+`hive-sh4re/src/permissions.rs` and to `Capability::ALL`. Implement the access
+check where hive-c0re enforces it (today: `set_nspawn_flags` in
+`hive-c0re/src/lifecycle/host_config.rs`).
 
 ## Async forms
 
diff --git a/docs/scheduler/coordinator.md b/docs/scheduler/coordinator.md
index 4634269f..09611c83 100644
--- a/docs/scheduler/coordinator.md
+++ b/docs/scheduler/coordinator.md
@@ -73,7 +73,7 @@ Cheap — no build slot:
 | `SetWanted`          | write the durable power intent (`wanted = Up`/`Offline`) as the head node of a power-op DAG. Takes the agent lease even though it's a store write, so the intent write and the tail `Reconcile` are atomic per-agent — two racing power ops can't clobber each other's intent before either reconciles                                                                                                     |
 | `EmitRebuilt`        | tail of a rebuild/perm-change — emit the agent's `Rebuilt` manager event (ok/fail per outcome, nothing on cancel). One node per agent _and_ per outcome                                                                                                                                                                                                                                                    |
 | `WriteDropin`        | `set_nspawn_flags` + `set_resource_limits` + daemon-reload                                                                                                                                                                                                                                                                                                                                                 |
-| `WritePermFile`      | commit `tool-groups.json` / `capabilities.json` (single git commit under `META_LOCK`) + emit the P3RM1SS10NS snapshots                                                                                                                                                                                                                                                                                     |
+| `WritePermFile`      | commit `tool-groups.json` (single git commit under `META_LOCK`) + emit the P3RM1SS10NS snapshot                                                                                                                                                                                                                                                                                                            |
 | `ForgeSweep`         | one-shot boot-time forge user/token sweep for every container (`forge::ensure_all`) as a first-class node, so it shows as real work on the dashboard instead of running invisibly in a bare `tokio::spawn`. Agentless                                                                                                                                                                                      |
 | `MatrixSweep`        | matrix user/space sweep (`matrix::ensure_all`): the boot-time instance, plus one every 30 min from a loop in `main.rs`. Holds `Resource::MatrixSweep` (capacity 1), so two passes never overlap; each tick queues its own pass, which waits for the resource if one is already live. Agentless                                                                                                             |
 | `KnowledgePull`      | `/knowledge` pull (`knowledge::pull`): at boot (commits that landed while `hive-c0re` was down), on the swarm knowledge-changed event, and hourly as a fallback. Holds `Resource::KnowledgeTree` (capacity 1), so two pulls never overlap on the working tree; each trigger queues its own pass, which waits for the resource if one is already live. Agentless                                            |
diff --git a/docs/tools/swarmctl-cli.md b/docs/tools/swarmctl-cli.md
index 3a54968c..8d3f6bb3 100644
--- a/docs/tools/swarmctl-cli.md
+++ b/docs/tools/swarmctl-cli.md
@@ -12,6 +12,10 @@ This document contains the help content for the `swarmctl` command-line program.
 * [`swarmctl agent inference get`↴](#swarmctl-agent-inference-get)
 * [`swarmctl agent inference set`↴](#swarmctl-agent-inference-set)
 * [`swarmctl agent inference clear`↴](#swarmctl-agent-inference-clear)
+* [`swarmctl agent permissions`↴](#swarmctl-agent-permissions)
+* [`swarmctl agent permissions get`↴](#swarmctl-agent-permissions-get)
+* [`swarmctl agent permissions set`↴](#swarmctl-agent-permissions-set)
+* [`swarmctl agent permissions clear`↴](#swarmctl-agent-permissions-clear)
 * [`swarmctl schedule`↴](#swarmctl-schedule)
 * [`swarmctl schedule add`↴](#swarmctl-schedule-add)
 * [`swarmctl schedule list`↴](#swarmctl-schedule-list)
@@ -66,6 +70,7 @@ Manage agents across the swarm
 * `create` — Queue creation of a new agent on a hive in this swarm
 * `mint-forge-token` — Check one agent's forge token, and mint it if it's missing or stale
 * `inference` — Show the model, effort, subagent models and subagent cap the swarm sets for an agent, or change the model, effort and subagent models
+* `permissions` — Show or change the capabilities the swarm grants an agent
 
 
 
@@ -198,6 +203,77 @@ Clear every field for an agent, so it goes back to its own values
 
 
 
+## `swarmctl agent permissions`
+
+Show or change the capabilities the swarm grants an agent.
+
+The hive running the agent reads them every 30 seconds and, when they change, restarts the agent after its current turn. Unset leaves the agent on its hive's default: none, except `ruth`, who gets `manage_root_agent`.
+
+**Usage:** `swarmctl agent permissions <COMMAND>`
+
+###### **Subcommands:**
+
+* `get` — Show the capabilities the swarm grants an agent, or that it sets none
+* `set` — Replace the capabilities the swarm grants an agent
+* `clear` — Clear the agent's permissions, so its hive's default applies again
+
+
+
+## `swarmctl agent permissions get`
+
+Show the capabilities the swarm grants an agent, or that it sets none
+
+**Usage:** `swarmctl agent permissions get [OPTIONS] <AGENT>`
+
+###### **Arguments:**
+
+* `<AGENT>` — Agent name
+
+###### **Options:**
+
+* `--controller-socket <PATH>` — swarm-controller's unix socket.
+
+   Supplied by the nix module that installs this binary, from the same `socketPath` option the daemon binds; falls back to `SWARM_CONTROLLER_SOCKET`.
+
+
+
+## `swarmctl agent permissions set`
+
+Replace the capabilities the swarm grants an agent
+
+**Usage:** `swarmctl agent permissions set [OPTIONS] --capabilities <CAPABILITIES> <AGENT>`
+
+###### **Arguments:**
+
+* `<AGENT>` — Agent name. A hive must declare it
+
+###### **Options:**
+
+* `--capabilities <CAPABILITIES>` — The capabilities to grant, comma-separated (`manage_root_agent`). `--capabilities ''` grants none
+* `--controller-socket <PATH>` — swarm-controller's unix socket.
+
+   Supplied by the nix module that installs this binary, from the same `socketPath` option the daemon binds; falls back to `SWARM_CONTROLLER_SOCKET`.
+
+
+
+## `swarmctl agent permissions clear`
+
+Clear the agent's permissions, so its hive's default applies again
+
+**Usage:** `swarmctl agent permissions clear [OPTIONS] <AGENT>`
+
+###### **Arguments:**
+
+* `<AGENT>` — Agent name
+
+###### **Options:**
+
+* `--controller-socket <PATH>` — swarm-controller's unix socket.
+
+   Supplied by the nix module that installs this binary, from the same `socketPath` option the daemon binds; falls back to `SWARM_CONTROLLER_SOCKET`.
+
+
+
 ## `swarmctl schedule`
 
 Manage scheduled prompts.
diff --git a/docs/web-ui/dashboard.md b/docs/web-ui/dashboard.md
index 6cfd43f1..38a3feac 100644
--- a/docs/web-ui/dashboard.md
+++ b/docs/web-ui/dashboard.md
@@ -77,8 +77,8 @@ H0M3 hub (served at
 `/core.html`), with minimal chrome: a `← home` back-link and a single pane,
 **K3PT ST4T3**. The page is its own esbuild bundle (`core.js`) that
 cold-loads `/api/state` and subscribes to `/api/dashboard/stream` for
-`tombstones_changed`, `capabilities_changed`, and `tool_groups_changed`
-(the latter two re-render the stale-perms sub-section when permission
+`tombstones_changed` and `tool_groups_changed`
+(the latter re-renders the stale-perms sub-section when permission
 data changes). (Rebuild queue and meta inputs are dashboard tabs — see
 below.)
 
@@ -104,11 +104,10 @@ container is all hive-c0re can actually prove. Fixing it properly
 needs a recorded destroy — **deferred** to the swarm-controller /
 snapshot-storage rework where the problem changes shape.
 
-*Stale permission entries*: agents with explicit capability or
-tool-group JSON entries but no live container — typically renamed or
-deleted agents whose entries persisted in `capabilities.json` /
-`tool-groups.json`. Loaded when the page opens;
-autorefreshes on `capabilities_changed` + `tool_groups_changed` SSE
+*Stale permission entries*: agents with explicit tool-group JSON
+entries but no live container — typically renamed or deleted agents
+whose entries persisted in `tool-groups.json`. Loaded when the page opens;
+autorefreshes on `tool_groups_changed` SSE
 events. Each ghost agent gets a `✕ clear perms` button
 (`DELETE /api/permissions/{name}`).
 
@@ -230,11 +229,13 @@ entry with that id pre-expanded.
 
 ## P3RM1SS10NS tab
 
-Per-agent permission configuration. Two sections, each rendered as a
-column-driven checkbox matrix: rows are agents, columns are the
-permission names fetched from hive-c0re. The column list is
-authoritative — adding a new tool-group or capability to hive-c0re
-requires no UI change; the new column appears automatically.
+Per-agent tool-group configuration, rendered as a column-driven
+checkbox matrix: rows are agents, columns are the group names fetched
+from hive-c0re. The column list is authoritative — adding a new
+tool-group to hive-c0re requires no UI change; the new column appears
+automatically. Capabilities aren't set here: they're the swarm's,
+set with `swarmctl agent permissions` (see
+`docs/process/conventions.md#capabilities`).
 
 The snapshot carries `agents` (the full manageable roster — live
 containers ∪ agents with an explicit entry) and `effective` (per-agent
@@ -248,29 +249,11 @@ strip those defaults. The `(default)` badge keys off absence from
 
 Fetches fire on tab activation (not page-load) to avoid unnecessary
 work when the operator never visits this tab. Live mutations from the
-rebuild-queue worker are also pushed via the `capabilities_changed` /
-`tool_groups_changed` SSE events (same payload shape as the GET
-endpoints), so an open P3RM1SS10NS tab reflects worker-applied changes
+rebuild-queue worker are also pushed via the `tool_groups_changed`
+SSE event (same payload shape as the GET endpoint), so an open P3RM1SS10NS tab reflects worker-applied changes
 without requiring navigation. Tab-activation re-fetches remain as a
 safety net for reconnect windows.
 
-**C4P4B1L1T13S** — per-agent capability grants. Capabilities unlock
-gated MCP tools and system-level access beyond the default agent
-surface. A saving POST queues a rebuild for the affected agent so the
-new `HIVE_CAPABILITIES` env var takes effect in the next session.
-
-The current capabilities are:
-
-| Name | Effect |
-|------|--------|
-| `manage_root_agent` | mounts every agent's state (rw) + config (ro) into this agent's container so it can manage/recover any agent |
-
-Each row is one agent. Columns are the capability names returned by
-`GET /api/capabilities` as `caps: Vec<String>`. Checking or unchecking
-boxes only stages the change in-browser; the browser writes nothing until
-the operator selects the page-level **save all** button (described below). Row
-values follow the `effective`/`assignments` rule described above.
-
 <!-- vale write-good.Passive = NO -->
 **T00L GR0UPS** — per-agent tool-group permissions. Tool groups are
 named buckets of MCP tools; each agent starts with a role default
@@ -291,10 +274,9 @@ The current tool groups are: `messaging`, `meta`, `inbox`, `lifecycle`,
 Claude's built-in `WebFetch` and `WebSearch` to `--tools` /
 `--allowedTools` for that agent session.
 
-Both tables share the same visual shape: `.cap-table-wrap` /
-`.tg-table-wrap` outer scroll container, `thead` with a label column
-(`.cap-agent-col` / `.tg-agent-col`) + one column per permission
-(`.cap-col` / `.tg-group-col`). Each tbody row is one agent: a name cell
+The table's shape: a `.tg-table-wrap` outer scroll container, `thead`
+with a label column (`.tg-agent-col`) + one column per group
+(`.tg-group-col`). Each tbody row is one agent: a name cell
 and its checkbox cells, where each checkbox carries `data-baseline` (its
 render-time state) and the row carries `data-agent` for dirty-tracking.
 
@@ -302,18 +284,15 @@ render-time state) and the row carries `data-agent` for dirty-tracking.
 buttons exist. A single page-level `.perm-save-bar` with a **save all (N
 agents)** button sits at the bottom of the pane, enabled only when some
 checkbox diverges from its baseline. Selecting it diffs every checkbox
-across *both* matrices and POSTs one batch to `POST /api/permissions` as
-`{ changes: [ { agent, tool_groups?, capabilities? } ] }` — the batch
-includes only the perm-types that actually changed for each agent (an
-omitted field leaves that file untouched; an included array fully
-replaces it).
-hive-c0re coalesces an agent's capabilities + tool-groups into a
-single rebuild, so changing both for one agent is one rebuild, not two.
+and POSTs one batch to `POST /api/permissions` as
+`{ changes: [ { agent, tool_groups } ] }` — the batch includes only the
+agents whose groups changed (an included array fully replaces the
+agent's groups).
 The batch is atomic: it validates every change first and on any error
 rejects the whole POST (`{error}`, nothing applied); a clean 200 (`ok`)
-flips the bar to a queued→rebuilding state and re-fetches both tables.
-Live `capabilities_changed` / `tool_groups_changed` events re-render the
-matrices unless the section has unsaved edits, so an in-progress edit set
+flips the bar to a queued→rebuilding state and re-fetches the table.
+Live `tool_groups_changed` events re-render the
+matrix unless it has unsaved edits, so an in-progress edit set
 isn't clobbered.
 
 ## ST4TS page (`/stats.html`)
@@ -767,31 +746,16 @@ below — some endpoints aren't in it yet.
   `/var/lib/hyperhive/meta/tool-groups.json` and queues a rebuild so
   `HIVE_TOOL_GROUPS` takes effect. Agent name validated;
   `guard_agent_name` applied.
-- `GET /api/capabilities` — returns `{ caps: Vec<String>,
-  assignments: BTreeMap<String, Vec<String>>,
-  descriptions: BTreeMap<String, String> }`. `caps` is the
-  ordered list of all known capability names; `assignments` is the
-  per-agent grant map (absent agents have no extra capabilities);
-  `descriptions` maps each capability name to a short human-readable
-  string surfaced as a column-header tooltip on hover.
-- `POST /api/capabilities/{agent}` — body `{ caps: ["name", …] }`.
-  Writes the capability set for `{agent}` to
-  `/var/lib/hyperhive/meta/capabilities.json` and queues a rebuild so
-  `HIVE_CAPABILITIES` takes effect. Agent name validated;
-  unknown capability strings rejected (400). `guard_agent_name`
-  applied.
 - `POST /api/permissions` — batch perm apply for the save-all
   permissions button. Body
-  `{ changes: [{ agent, tool_groups?: ["name", …], capabilities?: ["name", …] }] }`.
-  Sparse per agent: an omitted field leaves that perm-type untouched,
-  an empty array clears it, a populated array fully replaces it (same
-  replace semantics as the per-agent endpoints above). Each affected
-  agent gets ONE combined `PermChange` queue entry, so changing both
-  an agent's tool-groups and capabilities triggers a single rebuild,
-  not two. **Atomic**: the endpoint validates every change first (agent names via
-  `guard_agent_name`, group + capability names) and on any validation
-  error writes or enqueues nothing (non-2xx `{ error }`); it skips
-  rows with both fields omitted rather than erroring on them.
+  `{ changes: [{ agent, tool_groups?: ["name", …] }] }`.
+  An omitted `tool_groups` leaves the agent untouched, an empty array
+  clears it, a populated array fully replaces it (same replace
+  semantics as the per-agent endpoint above). Each affected agent gets
+  one `PermChange` queue entry. **Atomic**: the endpoint validates every
+  change first (agent names via `guard_agent_name`, group names) and on
+  any validation error writes or enqueues nothing (non-2xx `{ error }`);
+  it skips rows with the field omitted rather than erroring on them.
   Returns `200 "ok"` on
   success.
 - `POST /api/meta-update` — `nix flake update` the selected
@@ -841,13 +805,6 @@ payload):
   operator cancel / retry (`/api/system/reminders/*`), `cancel_loose_end`
   with Reminder kind, and the scheduler tick after each delivery
   batch (`reminder_scheduler`).
-- `capabilities_changed` (seq, caps: `Vec<str>`, descriptions: map,
-  assignments: `BTreeMap<String, Vec<String>>`) — full snapshot of
-  capability grants. Emitted from the rebuild-queue worker after a
-  `PermChange` / Capabilities entry commits the JSON file. Payload
-  matches `GET /api/capabilities` shape so the client can call
-  `renderCapabilities` directly. P3RM1SS10NS tab subscribes; activation
-  re-fetch still runs as a safety net.
 - `tool_groups_changed` (seq, groups: `Vec<str>`, descriptions: map,
   assignments: `BTreeMap<String, Vec<String>>`) — full snapshot of
   tool-group assignments. Emitted from the rebuild-queue worker after
diff --git a/frontend/packages/dashboard/src/core.html b/frontend/packages/dashboard/src/core.html
index 95ac4964..678def0e 100644
--- a/frontend/packages/dashboard/src/core.html
+++ b/frontend/packages/dashboard/src/core.html
@@ -27,8 +27,8 @@
 
     <main class="core-main">
       <!-- tombstones-section shows destroyed agents (purge button).
-         tombstones-stale-perms shows agents with explicit capability/
-         tool-group entries but no live container (typically renamed/deleted
+         tombstones-stale-perms shows agents with explicit tool-group
+         entries but no live container (typically renamed/deleted
          agents whose JSON entries persisted). Each stale entry gets a
          "✕ clear perms" button. -->
       <p class="meta">
diff --git a/frontend/packages/dashboard/src/core.js b/frontend/packages/dashboard/src/core.js
index 1e8919c1..459f2ff6 100644
--- a/frontend/packages/dashboard/src/core.js
+++ b/frontend/packages/dashboard/src/core.js
@@ -96,7 +96,7 @@ function renderTombstones(s) {
 }
 
 // ─── stale permission entries (K3PT ST4T3 pane sub-section) ──────────────
-// Agents with explicit capability / tool-group JSON entries but no live
+// Agents with explicit tool-group JSON entries but no live
 // container AND no kept-state tombstone (e.g. renamed agents like the old
 // "root" manager name). Ghost detection is server-side via
 // GET /api/permissions/stale so the client doesn't need to maintain a
@@ -114,7 +114,7 @@ function renderStalePerms(root, ghosts) {
     el(
       "p",
       { class: "meta" },
-      "agents with explicit capability or tool-group entries but no live container " +
+      "agents with explicit tool-group entries but no live container " +
         "or kept state (typically renamed or manually-deleted agents whose JSON entries persisted).",
     ),
   );
@@ -129,7 +129,7 @@ function renderStalePerms(root, ghosts) {
       {
         type: "button",
         class: "btn btn-destroy",
-        title: "remove explicit capability and tool-group entries for " + name,
+        title: "remove explicit tool-group entries for " + name,
       },
       "✕ clear perms",
     );
@@ -194,9 +194,6 @@ const SSE_HANDLERS = {
   },
   // Refresh the stale-perms sub-section when perm data changes (a ghost
   // was cleared, or perms were saved for an agent whose name collides).
-  capabilities_changed(_ev) {
-    if (stalePermsLoaded) fetchAndRenderStalePerms();
-  },
   tool_groups_changed(_ev) {
     if (stalePermsLoaded) fetchAndRenderStalePerms();
   },
@@ -234,8 +231,7 @@ async function init() {
   // (subscription discipline, part 1 of the dashboard-event-stream-
   // split issue).
   const es = openStream(
-    "/api/dashboard/stream?kinds=tombstones_changed," +
-      "capabilities_changed,tool_groups_changed",
+    "/api/dashboard/stream?kinds=tombstones_changed,tool_groups_changed",
   );
   if (es) {
     es.onmessage = (e) => {
diff --git a/frontend/packages/dashboard/src/dashboard.css b/frontend/packages/dashboard/src/dashboard.css
index c1ecf3c6..94713cbe 100644
--- a/frontend/packages/dashboard/src/dashboard.css
+++ b/frontend/packages/dashboard/src/dashboard.css
@@ -566,57 +566,9 @@ footer .banner-thin {
   margin-bottom: 0.8em;
 }
 
-/* ─── capabilities (permissions) table ─────────────────────────────
-   Agents × capabilities matrix in the P3RM1SS10NS tab. Same layout
-   as the tool-groups table below. Horizontally scrollable on narrow
-   viewports. */
-.cap-table-wrap {
-  overflow-x: auto;
-  margin-top: 0.5em;
-}
-.cap-table {
-  border-collapse: collapse;
-  font-size: 0.82em;
-  min-width: 100%;
-}
-.cap-table th,
-.cap-table td {
-  padding: 0.35em 0.6em;
-  border: 1px solid var(--border);
-  text-align: center;
-  vertical-align: middle;
-}
-.cap-table thead th {
-  background: var(--bg-elev);
-  color: var(--muted);
-  letter-spacing: 0.05em;
-  white-space: nowrap;
-}
-.cap-agent-col {
-  text-align: left !important;
-  min-width: 8em;
-}
-.cap-col {
-  min-width: 5em;
-}
-.cap-agent-name {
-  color: var(--fg);
-  font-weight: 600;
-}
-.cap-cb {
-  cursor: pointer;
-  width: 1em;
-  height: 1em;
-  accent-color: var(--purple);
-}
-.cap-row:hover td {
-  background: var(--bg-elev);
-}
-
 /* ─── tool-groups (permissions) table ──────────────────────────────
-   Agents × tool-groups matrix in the P3RM1SS10NS tab. Same layout
-   as the capabilities table above. Horizontally scrollable on narrow
-   viewports. */
+   Agents × tool-groups matrix in the P3RM1SS10NS tab. Horizontally
+   scrollable on narrow viewports. */
 .tg-table-wrap {
   overflow-x: auto;
   margin-top: 0.5em;
diff --git a/frontend/packages/dashboard/src/dashboard.html b/frontend/packages/dashboard/src/dashboard.html
index fb889a80..2fbd405f 100644
--- a/frontend/packages/dashboard/src/dashboard.html
+++ b/frontend/packages/dashboard/src/dashboard.html
@@ -71,8 +71,8 @@
         >
           <span class="tab-label">◆ BUILD L0GS ◆</span>
         </a>
-        <!-- P3RM1SS10NS: per-agent capability grants and tool-group
-           assignments. Both tables are fetched on tab activation. -->
+        <!-- P3RM1SS10NS: per-agent tool-group
+           assignments, fetched on tab activation. -->
         <a
           class="tab"
           id="tab-permissions"
@@ -205,13 +205,12 @@
           <div id="build-list"><p class="meta">loading…</p></div>
         </section>
 
-        <!-- P3RM1SS10NS: per-agent capability grants + tool-group
-         assignments. Both tables are column-driven from the backend
-         (GET /api/capabilities, GET /api/tool-groups) so new entries
-         require no UI change. Edits stage in-browser across both
-         matrices; the single page-level "save all" button POSTs every
-         changed agent to /api/permissions in one atomic batch (caps +
-         groups coalesce into one rebuild per agent). -->
+        <!-- P3RM1SS10NS: per-agent tool-group assignments, column-driven
+         from the backend (GET /api/tool-groups) so new groups require no
+         UI change. Edits stage in-browser; the page-level "save all"
+         button POSTs every changed agent to /api/permissions in one
+         atomic batch (one rebuild per agent). Capabilities are the
+         swarm's (`swarmctl agent permissions`), not set here. -->
         <section
           class="tab-pane"
           id="tab-pane-permissions"
@@ -220,25 +219,11 @@
           role="tabpanel"
           aria-labelledby="tab-permissions"
         >
-          <h2>◆ C4P4B1L1T13S ◆</h2>
-          <div class="divider">
-            ══════════════════════════════════════════════════════════════
-          </div>
-          <p class="meta">
-            per-agent capability grants. capabilities unlock gated MCP tools and
-            system access. toggle any cells across both tables, then hit "save
-            all" at the bottom — each changed agent rebuilds exactly once.
-          </p>
-          <div id="capabilities-section">
-            <p class="meta">loading…</p>
-          </div>
-
           <!-- T00L GR0UPS: per-agent tool-group permission matrix. Rows = agents,
            cols = tool groups fetched from GET /api/tool-groups.
            Checking / unchecking is staged in the UI; the single page-level
            "save all" button POSTs every changed agent to
-           /api/permissions in one atomic batch, coalescing caps + groups
-           per agent into one rebuild each.
+           /api/permissions in one atomic batch, one rebuild each.
            Absent agents default to the role default (shown in parens). -->
           <h2>◆ T00L GR0UPS ◆</h2>
           <div class="divider">
@@ -255,10 +240,9 @@
           </div>
 
           <!-- Page-level save bar (save-all issue): one button for the whole
-           permissions page. Staged checkbox edits across BOTH matrices are
-           diffed against their render-time baselines and POSTed together to
-           /api/permissions; caps + groups for one agent coalesce into a
-           single rebuild. Disabled until something is dirty. The atomic
+           permissions page. Staged checkbox edits are diffed against their
+           render-time baselines and POSTed together to /api/permissions, one
+           rebuild per agent. Disabled until something is dirty. The atomic
            batch either fully lands (queued ✓ → rebuilding) or fully fails
            (error note, nothing applied). -->
           <div class="perm-save-bar">
diff --git a/frontend/packages/dashboard/src/permissions.js b/frontend/packages/dashboard/src/permissions.js
index e4aa0794..367079a3 100644
--- a/frontend/packages/dashboard/src/permissions.js
+++ b/frontend/packages/dashboard/src/permissions.js
@@ -1,11 +1,11 @@
-// Dashboard P3RM1SS10NS tab — the per-agent capabilities + tool-groups
-// matrices, with a single page-level "save all" button.
+// Dashboard P3RM1SS10NS tab — the per-agent tool-groups matrix, with a
+// page-level "save all" button. Capabilities are not set here: they are
+// the swarm's (`swarmctl agent permissions`).
 //
-// Both tables are fetched on tab activation (`GET /api/capabilities`,
-// `GET /api/tool-groups`) and after each save. Columns (caps / groups)
-// come from the backend so the UI needs no change when a new one is
-// added. Live updates arrive via the `capabilities_changed` /
-// `tool_groups_changed` dashboard events (fired after the rebuild-queue
+// The table is fetched on tab activation (`GET /api/tool-groups`) and
+// after each save. Columns come from the backend so the UI needs no
+// change when a new group is added. Live updates arrive via the
+// `tool_groups_changed` dashboard event (fired after the rebuild-queue
 // worker commits the perm JSON file), wired into the entry's mutation
 // dispatch table.
 //
@@ -15,11 +15,9 @@
 // assignments map.
 //
 // Editing model: checkboxes are diffed against a `data-baseline` captured at
-// render time; only changed perm-types per agent are POSTed to
-// `POST /api/permissions` as `{ changes: [{agent, tool_groups?, capabilities?}] }`.
-// Omitted field = leave untouched; included array = full replace. The backend
-// coalesces caps+groups per agent into ONE queue entry (one rebuild, no
-// double-rebuild). Batch is atomic — saved→rebuilding only fires on a clean 200.
+// render time; only changed agents are POSTed to `POST /api/permissions` as
+// `{ changes: [{agent, tool_groups}] }` (a full replace per agent). Batch is
+// atomic — saved→rebuilding only fires on a clean 200.
 
 import { $ } from "./common.js";
 import { el } from "@hive/shared/dom.js";
@@ -32,13 +30,6 @@ import { asyncBtn } from "@hive/shared/forms.js";
 // is worse than a brief staleness window. The tab-activation re-fetch
 // and the post-save re-fetch are the recovery paths; both clear dirty.
 
-export function applyCapabilitiesChanged(ev) {
-  const root = $("capabilities-section");
-  if (!root) return;
-  if (root.contains(document.activeElement)) return;
-  if (sectionHasDirty(root)) return;
-  renderCapabilities(root, ev);
-}
 export function applyToolGroupsChanged(ev) {
   const root = $("tool-groups-section");
   if (!root) return;
@@ -56,121 +47,6 @@ function sectionHasDirty(root) {
   return false;
 }
 
-export async function fetchAndRenderCapabilities() {
-  const root = $("capabilities-section");
-  if (!root) return;
-  root.replaceChildren();
-  root.append(el("p", { class: "meta" }, "loading…"));
-  try {
-    const resp = await fetch("/api/capabilities");
-    if (!resp.ok) throw new Error("http " + resp.status);
-    const data = await resp.json();
-    renderCapabilities(root, data);
-  } catch (err) {
-    root.replaceChildren();
-    root.append(el("p", { class: "meta" }, "fetch failed: " + err));
-  }
-}
-
-function renderCapabilities(root, data) {
-  root.replaceChildren();
-  const { caps, descriptions = {}, assignments, effective = {} } = data;
-  if (!caps || !caps.length) {
-    root.append(el("p", { class: "meta" }, "(no capabilities defined)"));
-    updateSaveBar();
-    return;
-  }
-
-  // Agent rows: prefer the backend roster (every manageable agent,
-  // default-perms included). Fall back to the live-container ∪ explicit
-  // union for older payloads that don't carry `agents`.
-  const agentNames =
-    data.agents && data.agents.length
-      ? [...data.agents]
-      : [
-          ...new Set([
-            ...Array.from(containersState.keys()),
-            ...Object.keys(assignments),
-          ]),
-        ].sort();
-
-  if (!agentNames.length) {
-    root.append(el("p", { class: "meta" }, "(no agents)"));
-    updateSaveBar();
-    return;
-  }
-
-  const wrap = el("div", { class: "cap-table-wrap" });
-  const table = el("table", { class: "cap-table" });
-
-  // Header row.
-  const thead = el("thead");
-  const hrow = el("tr");
-  hrow.append(el("th", { class: "cap-agent-col" }, "agent"));
-  for (const c of caps) {
-    hrow.append(el("th", { class: "cap-col", title: descriptions[c] || c }, c));
-  }
-  thead.append(hrow);
-  table.append(thead);
-
-  const tbody = el("tbody");
-  for (const name of agentNames) {
-    // Effective caps (explicit-or-default) drive the checkboxes so
-    // default-perms agents show their real grants, not blank.
-    const assigned = effective[name] || assignments[name] || [];
-    // `containersState` is keyed from `nixos-container list`, which
-    // includes stopped-but-configured containers — so a temporarily-stopped
-    // agent is NOT stale. Only destroyed/renamed agents are absent here.
-    const isStale = !containersState.has(name);
-    const tr = el("tr", {
-      class: "cap-row" + (isStale ? " perm-row-stale" : ""),
-      "data-agent": name,
-    });
-
-    // Agent name cell.
-    const nameTd = el("td", { class: "cap-agent-col" });
-    nameTd.append(el("span", { class: "cap-agent-name" }, name));
-    if (isStale) {
-      nameTd.append(el("span", { class: "perm-stale-label" }, "(not running)"));
-      const removeBtn = el(
-        "button",
-        {
-          type: "button",
-          class: "perm-remove-btn",
-          title: "remove stale permission entries for " + name,
-        },
-        "✕ remove",
-      );
-      removeBtn.addEventListener("click", () => clearStaleAgent(name, root));
-      nameTd.append(removeBtn);
-    }
-    tr.append(nameTd);
-
-    // One checkbox per capability.
-    for (const c of caps) {
-      const checked = assigned.includes(c);
-      const td = el("td", { class: "cap-col" });
-      const cb = el("input", {
-        type: "checkbox",
-        class: "cap-cb",
-        "data-cap": c,
-        "data-baseline": checked ? "1" : "0",
-        "aria-label": c,
-      });
-      cb.checked = checked;
-      cb.addEventListener("change", onCellToggle);
-      td.append(cb);
-      tr.append(td);
-    }
-
-    tbody.append(tr);
-  }
-  table.append(tbody);
-  wrap.append(table);
-  root.append(wrap);
-  updateSaveBar();
-}
-
 export async function fetchAndRenderToolGroups() {
   const root = $("tool-groups-section");
   if (!root) return;
@@ -303,20 +179,11 @@ function onCellToggle() {
   updateSaveBar();
 }
 
-// Collect the sparse change-set: one entry per agent that has at least
-// one perm-type diverging from baseline. Only the changed perm-type(s)
-// are included so untouched defaults stay defaults (an included array is
-// a full replacement → omitting it leaves that file alone).
+// Collect the sparse change-set: one entry per agent whose groups diverge
+// from baseline, so untouched defaults stay defaults (an included array is
+// a full replacement).
 function collectChanges() {
-  const byAgent = new Map(); // agent -> { capabilities?, tool_groups? }
-  collectSection(
-    $("capabilities-section"),
-    ".cap-row",
-    ".cap-cb",
-    "cap",
-    "capabilities",
-    byAgent,
-  );
+  const byAgent = new Map(); // agent -> { tool_groups }
   collectSection(
     $("tool-groups-section"),
     ".tg-row",
@@ -363,7 +230,7 @@ function updateSaveBar() {
 // Remove all explicit permission entries for a stale (non-running)
 // agent. Calls DELETE /api/permissions/{agent}, which bypasses the
 // roster guard so the stale entries can be cleaned up even though the
-// agent isn't in the live container list. Re-fetches both tables after
+// agent isn't in the live container list. Re-fetches the table after
 // the delete so the row disappears immediately.
 async function clearStaleAgent(name, sectionRoot) {
   const btn = sectionRoot
@@ -381,11 +248,8 @@ async function clearStaleAgent(name, sectionRoot) {
         setSaveNote("failed to remove " + name + ": " + text, true);
         return;
       }
-      // Re-fetch both sections so the stale row disappears.
-      await Promise.all([
-        fetchAndRenderCapabilities(),
-        fetchAndRenderToolGroups(),
-      ]);
+      // Re-fetch so the stale row disappears.
+      await fetchAndRenderToolGroups();
     } catch (err) {
       setSaveNote("failed to remove " + name + ": " + err, true);
     }
@@ -437,14 +301,13 @@ async function saveAll() {
       return;
     }
     // Atomic 200: the whole batch landed. Show a rebuilding hint and
-    // re-fetch both tables once the queue worker has committed — that
+    // re-fetch the table once the queue worker has committed — that
     // resets baselines (dirty clears) and the bar disables itself.
     const n = changes.length;
     btn.textContent = "queued ✓";
     setSaveNote(`rebuilding ${n} agent${n === 1 ? "" : "s"}…`, false);
     setTimeout(() => {
       btn.dataset.busy = "";
-      fetchAndRenderCapabilities();
       fetchAndRenderToolGroups();
     }, 900);
   } catch (err) {
diff --git a/frontend/packages/dashboard/src/state.js b/frontend/packages/dashboard/src/state.js
index 0b3b7a9e..09df4747 100644
--- a/frontend/packages/dashboard/src/state.js
+++ b/frontend/packages/dashboard/src/state.js
@@ -2,7 +2,7 @@
 // dashboard SPA: the live agent roster.
 //
 // The roster is read by nearly every tab (the container tree, the
-// capabilities + tool-group matrices, and the selection bar), so
+// tool-group matrix, and the selection bar), so
 // it lives here as a single source of truth that each domain module
 // imports rather than threading through call signatures. Everything
 // else in the dashboard is domain-local and lives with its own module.
diff --git a/frontend/packages/dashboard/src/system-sections.css b/frontend/packages/dashboard/src/system-sections.css
index 1f862cab..c8da583a 100644
--- a/frontend/packages/dashboard/src/system-sections.css
+++ b/frontend/packages/dashboard/src/system-sections.css
@@ -201,9 +201,9 @@
    sub-step label` — the rule has had no JS referents since then). */
 
 /* ─── K3PT ST4T3 › stale permission entries sub-section ────────────────────
-   Agents with explicit capability / tool-group entries in the JSON but no
-   live container (renamed or manually-deleted agents). Lazy-loaded on tab
-   activation, refreshed on capabilities_changed / tool_groups_changed. */
+   Agents with explicit tool-group entries in the JSON but no live
+   container (renamed or manually-deleted agents). Lazy-loaded on tab
+   activation, refreshed on tool_groups_changed. */
 .tombstones-stale-heading {
   margin: 1.2em 0 0.3em;
   color: var(--muted);
diff --git a/frontend/packages/dashboard/src/tabs.js b/frontend/packages/dashboard/src/tabs.js
index adef8516..df62a678 100644
--- a/frontend/packages/dashboard/src/tabs.js
+++ b/frontend/packages/dashboard/src/tabs.js
@@ -17,9 +17,7 @@ import { bindAsyncForms } from "@hive/shared/forms.js";
 import { createTabStrip } from "@hive/shared/tabs.js";
 import { containersState, syncContainersFromSnapshot } from "./state.js";
 import {
-  applyCapabilitiesChanged,
   applyToolGroupsChanged,
-  fetchAndRenderCapabilities,
   fetchAndRenderToolGroups,
   initPermissions,
 } from "./permissions.js";
@@ -64,7 +62,6 @@ window.marked = marked;
   const MANAGED_SECTION_IDS = [
     "containers-section",
     "inbox-section",
-    "capabilities-section",
     "tool-groups-section",
   ];
   // <details> sections that should survive a refresh need a stable
@@ -210,7 +207,6 @@ window.marked = marked;
     },
     meta_inputs_changed: onMetaInputsChanged,
     meta_update_running: onMetaUpdateRunning,
-    capabilities_changed: applyCapabilitiesChanged,
     tool_groups_changed: applyToolGroupsChanged,
   };
   (function bindDashboardStream() {
@@ -228,7 +224,7 @@ window.marked = marked;
         "transient_set,transient_cleared," +
         "container_state_changed,container_removed,rebuild_queue_changed," +
         "meta_inputs_changed,meta_update_running," +
-        "capabilities_changed,tool_groups_changed",
+        "tool_groups_changed",
     );
     es.onmessage = (e) => {
       let ev;
@@ -278,13 +274,12 @@ window.marked = marked;
     // visibility (bar only belongs on 4G3NTS where agent cards live).
     document.body.dataset.activeTab = target;
     renderSelectionBar(Array.from(containersState.values()));
-    // Permissions tables: SSE covers worker-applied changes
-    // (capabilities_changed / tool_groups_changed); re-fetch on
+    // Permissions table: SSE covers worker-applied changes
+    // (tool_groups_changed); re-fetch on
     // activation as a safety net for any gap between SSE events and
     // the cold-load snapshot.
     if (target === "permissions") {
       initPermissions();
-      fetchAndRenderCapabilities();
       fetchAndRenderToolGroups();
     }
     onBuildsTabShown(target);
diff --git a/hive-c0re/Cargo.toml b/hive-c0re/Cargo.toml
index 9e8d0f0c..59ee8731 100644
--- a/hive-c0re/Cargo.toml
+++ b/hive-c0re/Cargo.toml
@@ -57,6 +57,7 @@ swarm-queue-client = { workspace = true, features = [
   "kv",
   "notices",
   "agent-limits",
+  "agent-permissions",
 ] }
 swarm-secret-client.workspace = true
 tokio.workspace = true
diff --git a/hive-c0re/src/agent_config/capabilities.rs b/hive-c0re/src/agent_config/capabilities.rs
index c520b0e6..41b99fb2 100644
--- a/hive-c0re/src/agent_config/capabilities.rs
+++ b/hive-c0re/src/agent_config/capabilities.rs
@@ -1,262 +1,109 @@
-//! Per-agent capability configuration. Stored at
-//! `/var/lib/hyperhive/meta/capabilities.json` alongside `topology.json`
-//! and `tool-groups.json`.
+//! The capabilities the swarm grants this hive's agents, as last read from the
+//! `agent-permissions` bucket by `workers::swarm_permissions`. Stored at
+//! `/var/lib/hyperhive/swarm-agent-permissions.json`, outside the meta repo:
+//! the swarm is the record, this file is the hive's last read of it.
 //!
-//! Format: a JSON object mapping agent name to an array of
-//! `hive_sh4re::permissions::Capability` `snake_case` strings:
+//! Format: agent name → the stored
+//! [`swarm_queue_client::agent_permissions::AgentPermissions`]. An agent with
+//! no entry, or an entry with `capabilities` unset, holds its default
+//! ([`default_for`]); `"capabilities": []` holds none.
 //!
-//! ```json
-//! {
-//!   "atlas": ["manage_root_agent"]
-//! }
-//! ```
-//!
-//! An absent entry (or an absent file) means "no extra capabilities".
-//! `render_flake` in `meta.rs` reads this file and injects
-//! `HIVE_CAPABILITIES` into each agent's systemd service env; absent
-//! entries emit no env var so agents without capabilities don't trigger
-//! a spurious rebuild.
-//!
-//! Write path: `set_caps` is called from the dashboard action handler
-//! that the operator uses to grant/revoke capabilities per agent.
+//! Read path: `lifecycle::host_config::set_nspawn_flags`, on every container
+//! start, which is why a change takes a restart.
 
-use hive_sh4re::permissions::Capability;
 use std::collections::BTreeMap;
 use std::path::{Path, PathBuf};
 
-const CAPABILITIES_FILE: &str = "capabilities.json";
+use hive_sh4re::permissions::Capability;
+use swarm_queue_client::agent_permissions::AgentPermissions;
+
+const FILE: &str = "swarm-agent-permissions.json";
+
+pub type Stored = BTreeMap<String, AgentPermissions>;
 
 #[must_use]
-pub fn capabilities_path() -> PathBuf {
-    crate::paths::meta_root().join(CAPABILITIES_FILE)
+pub fn path() -> PathBuf {
+    crate::paths::state_root().join(FILE)
 }
 
-/// True if `name` matches a recognised [`Capability`], case-insensitive
-/// (same comparison [`has_cap`] uses).
+/// Read the stored permissions. An absent file is the empty map: every agent
+/// on its default. A file that exists but can't be read or parsed is an error.
+pub fn read_at(path: &Path) -> std::io::Result<Stored> {
+    super::read_map(path)
+}
+
+/// Persist the permissions atomically.
+pub fn write_at(path: &Path, stored: &Stored) -> std::io::Result<()> {
+    super::write_map(path, stored)
+}
+
+/// The capabilities an agent holds when the swarm sets none: `ruth` recovers
+/// the other agents, every other agent holds nothing.
 #[must_use]
-fn is_known(name: &str) -> bool {
-    Capability::ALL
-        .iter()
-        .any(|cap| name.eq_ignore_ascii_case(<&str>::from(*cap)))
-}
-
-/// Drop any capability string in `map` that isn't a recognised
-/// [`Capability`], warning per dropped string so a typo'd or stale
-/// grant doesn't look like it worked. An agent left with no
-/// capabilities is removed entirely, matching [`set_caps`]'s
-/// empty-vec behaviour. Returns whether anything was dropped.
-fn prune_unknown(map: &mut BTreeMap<String, Vec<String>>) -> bool {
-    let mut dropped_any = false;
-    map.retain(|agent, caps| {
-        caps.retain(|cap| {
-            let known = is_known(cap);
-            if !known {
-                dropped_any = true;
-                tracing::warn!(agent, capability = %cap, "dropping unrecognised capability");
-            }
-            known
-        });
-        !caps.is_empty()
-    });
-    dropped_any
-}
-
-/// Read the per-agent capability map. An absent file is the empty map —
-/// callers treat a missing entry as "no extra capabilities". A file
-/// that exists but can't be read or parsed is an error. Any entry that
-/// isn't a recognised [`Capability`] is dropped (with a `warn!`) from
-/// what's returned — a stale or typo'd name is never honoured — but
-/// `read` itself never writes. Unknown names stay on disk until the next
-/// write: `set_caps` always rewrites the file, `remove_agent` only when
-/// the agent had an entry.
-pub fn read() -> std::io::Result<BTreeMap<String, Vec<String>>> {
-    read_from(&capabilities_path())
-}
-
-fn read_from(path: &Path) -> std::io::Result<BTreeMap<String, Vec<String>>> {
-    let mut map = super::read_map(path)?;
-    prune_unknown(&mut map);
-    Ok(map)
-}
-
-/// Look up the configured capabilities for one agent in the file at
-/// `path` (normally [`capabilities_path`]). Returns an empty vec when
-/// the agent has no entry. Errors as [`read`] does.
-pub fn caps_for(path: &Path, name: &str) -> std::io::Result<Vec<String>> {
-    Ok(read_from(path)?.get(name).cloned().unwrap_or_default())
-}
-
-/// Check whether an agent holds a specific capability, per the file at
-/// `path` (normally [`capabilities_path`]). Errors as [`read`] does.
-pub fn has_cap(
-    path: &Path,
-    name: &str,
-    cap: hive_sh4re::permissions::Capability,
-) -> std::io::Result<bool> {
-    Ok(caps_for(path, name)?
-        .iter()
-        .any(|s| s.eq_ignore_ascii_case(<&str>::from(cap))))
-}
-
-/// Filter `caps` down to recognised [`Capability`] names (warning per
-/// drop) and apply the result to `current` for `name`: an empty
-/// result — whether `caps` started empty or every name in it was
-/// unknown — removes the entry, matching `read`'s empty-entry
-/// convention. Split out from [`set_caps`] so the decision logic is
-/// testable without touching the real capabilities file.
-fn apply_known_caps(current: &mut BTreeMap<String, Vec<String>>, name: &str, caps: &[String]) {
-    let known: Vec<String> = caps
-        .iter()
-        .filter(|cap| {
-            let ok = is_known(cap);
-            if !ok {
-                tracing::warn!(agent = name, capability = %cap, "dropping unrecognised capability");
-            }
-            ok
-        })
-        .cloned()
-        .collect();
-    if known.is_empty() {
-        current.remove(name);
+pub fn default_for(agent: &str) -> Vec<Capability> {
+    if agent == crate::lifecycle::MANAGER_NAME {
+        vec![Capability::ManageRootAgent]
     } else {
-        current.insert(name.to_owned(), known);
+        Vec::new()
     }
 }
 
-/// Set the capabilities for one agent and persist the map. Any name
-/// that isn't a recognised [`Capability`] is dropped (with a `warn!`)
-/// rather than written — an unknown grant should never look like it
-/// took effect. An empty `caps` vec, or one that becomes empty after
-/// dropping unknown names, removes the entry (agent has no
-/// capabilities). Fails without writing if the existing file can't be
-/// read or parsed.
-pub fn set_caps(name: &str, caps: &[String]) -> std::io::Result<()> {
-    set_caps_at(&capabilities_path(), name, caps)
+/// The capabilities `agent` holds under `record`.
+#[must_use]
+pub fn effective(agent: &str, record: Option<&AgentPermissions>) -> Vec<Capability> {
+    record
+        .and_then(|r| r.capabilities.clone())
+        .unwrap_or_else(|| default_for(agent))
 }
 
-fn set_caps_at(path: &Path, name: &str, caps: &[String]) -> std::io::Result<()> {
-    let mut current = read_from(path)?;
-    apply_known_caps(&mut current, name, caps);
-    super::write_map(path, &current)
-}
-
-/// Remove an agent from the capability map entirely. Called by
-/// `meta::sync_agents` when an agent is deprovisioned so stale entries
-/// don't accumulate. No-op if the agent has no entry. Fails without
-/// writing if the existing file can't be read or parsed.
-pub fn remove_agent(name: &str) -> std::io::Result<()> {
-    remove_agent_at(&capabilities_path(), name)
-}
-
-fn remove_agent_at(path: &Path, name: &str) -> std::io::Result<()> {
-    let mut current = read_from(path)?;
-    if current.remove(name).is_some() {
-        super::write_map(path, &current)?;
-    }
-    Ok(())
+/// Whether `agent` holds `cap` per the file at `path` (normally [`path`]).
+///
+/// # Errors
+///
+/// As [`read_at`].
+pub fn holds(path: &Path, agent: &str, cap: Capability) -> std::io::Result<bool> {
+    Ok(effective(agent, read_at(path)?.get(agent)).contains(&cap))
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
 
-    // `read`/`set_caps`/`remove_agent` call `crate::paths::meta_root`,
-    // which is hardcoded to `/var/lib/hyperhive` (no test override) — so
-    // these tests pin the pure decision logic (`prune_unknown`,
-    // `apply_known_caps`) against in-memory maps, and the file handling
-    // through the path-taking `*_at` / `read_from` variants.
-
-    const TRUNCATED: &str = "{\n  \"atlas\": [\"manage_root_ag";
-
-    fn corrupt_file() -> (tempfile::TempDir, std::path::PathBuf) {
-        let dir = tempfile::tempdir().expect("tempdir");
-        let path = dir.path().join(CAPABILITIES_FILE);
-        std::fs::write(&path, TRUNCATED).expect("seed");
-        (dir, path)
+    fn record(json: &str) -> AgentPermissions {
+        serde_json::from_str(json).expect("decodes")
     }
 
     #[test]
-    fn read_of_a_corrupt_file_is_an_error() {
-        let (_dir, path) = corrupt_file();
-        read_from(&path).expect_err("truncated JSON must not read as a map");
-    }
-
-    #[test]
-    fn set_caps_leaves_a_corrupt_file_untouched() {
-        let (_dir, path) = corrupt_file();
-        let err = set_caps_at(&path, "ruth", &["manage_root_agent".to_owned()])
-            .expect_err("a corrupt file must not be overwritten");
-        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
-        assert_eq!(std::fs::read(&path).expect("read"), TRUNCATED.as_bytes());
-    }
-
-    #[test]
-    fn remove_agent_leaves_a_corrupt_file_untouched() {
-        let (_dir, path) = corrupt_file();
-        let err =
-            remove_agent_at(&path, "atlas").expect_err("a corrupt file must not be overwritten");
-        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
-        assert_eq!(std::fs::read(&path).expect("read"), TRUNCATED.as_bytes());
-    }
-
-    #[test]
-    fn set_caps_round_trips_through_a_missing_file() {
-        let dir = tempfile::tempdir().expect("tempdir");
-        let path = dir.path().join(CAPABILITIES_FILE);
-        assert!(read_from(&path).expect("missing file is empty").is_empty());
-        set_caps_at(&path, "ruth", &["manage_root_agent".to_owned()]).expect("set");
-        let map = read_from(&path).expect("read");
-        assert_eq!(map["ruth"], vec!["manage_root_agent".to_owned()]);
-    }
-
-    #[test]
-    fn prune_unknown_keeps_known_name() {
-        let mut map = BTreeMap::from([("atlas".to_owned(), vec!["manage_root_agent".to_owned()])]);
-        assert!(!prune_unknown(&mut map));
-        assert_eq!(map["atlas"], vec!["manage_root_agent".to_owned()]);
-    }
-
-    #[test]
-    fn prune_unknown_drops_unrecognised_name() {
-        let mut map = BTreeMap::from([("atlas".to_owned(), vec!["fly_to_the_moon".to_owned()])]);
-        assert!(prune_unknown(&mut map));
-        assert!(!map.contains_key("atlas"));
-    }
-
-    #[test]
-    fn prune_unknown_keeps_known_and_drops_unknown_in_same_entry() {
-        let mut map = BTreeMap::from([(
-            "atlas".to_owned(),
-            vec!["manage_root_agent".to_owned(), "fly_to_the_moon".to_owned()],
-        )]);
-        assert!(prune_unknown(&mut map));
-        assert_eq!(map["atlas"], vec!["manage_root_agent".to_owned()]);
-    }
-
-    #[test]
-    fn apply_known_caps_removes_entry_on_empty_input() {
-        let mut current =
-            BTreeMap::from([("atlas".to_owned(), vec!["manage_root_agent".to_owned()])]);
-        apply_known_caps(&mut current, "atlas", &[]);
-        assert!(!current.contains_key("atlas"));
-    }
-
-    #[test]
-    fn apply_known_caps_removes_entry_when_only_unknown_names_given() {
-        let mut current =
-            BTreeMap::from([("atlas".to_owned(), vec!["manage_root_agent".to_owned()])]);
-        apply_known_caps(&mut current, "atlas", &["fly_to_the_moon".to_owned()]);
-        assert!(!current.contains_key("atlas"));
-    }
-
-    #[test]
-    fn apply_known_caps_keeps_only_the_known_name() {
-        let mut current = BTreeMap::new();
-        apply_known_caps(
-            &mut current,
-            "atlas",
-            &["manage_root_agent".to_owned(), "fly_to_the_moon".to_owned()],
+    fn unset_is_the_default_and_an_empty_list_is_none() {
+        assert_eq!(effective("ruth", None), vec![Capability::ManageRootAgent]);
+        assert_eq!(
+            effective("ruth", Some(&record("{}"))),
+            vec![Capability::ManageRootAgent]
         );
-        assert_eq!(current["atlas"], vec!["manage_root_agent".to_owned()]);
+        assert!(effective("ruth", Some(&record(r#"{"capabilities":[]}"#))).is_empty());
+        assert!(effective("atlas", None).is_empty());
+        assert_eq!(
+            effective(
+                "atlas",
+                Some(&record(r#"{"capabilities":["manage_root_agent"]}"#))
+            ),
+            vec![Capability::ManageRootAgent]
+        );
+    }
+
+    #[test]
+    fn a_missing_file_holds_the_defaults() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let path = dir.path().join(FILE);
+        assert!(holds(&path, "ruth", Capability::ManageRootAgent).expect("missing is empty"));
+        assert!(!holds(&path, "atlas", Capability::ManageRootAgent).expect("missing is empty"));
+    }
+
+    #[test]
+    fn a_corrupt_file_is_an_error() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let path = dir.path().join(FILE);
+        std::fs::write(&path, "{\n  \"ruth\": {\"capabilities\": [\"manage_root_ag").expect("seed");
+        holds(&path, "ruth", Capability::ManageRootAgent).expect_err("truncated JSON");
     }
 }
diff --git a/hive-c0re/src/agent_config/mod.rs b/hive-c0re/src/agent_config/mod.rs
index 04dee8f5..af5a36eb 100644
--- a/hive-c0re/src/agent_config/mod.rs
+++ b/hive-c0re/src/agent_config/mod.rs
@@ -1,6 +1,6 @@
-//! Per-agent configuration registries: tool groups, capabilities,
-//! resource limits, topology (all JSON files under
-//! `/var/lib/hyperhive/meta/`) and the shared wire-protocol size
+//! Per-agent configuration registries: tool groups, resource limits,
+//! topology (JSON files under `/var/lib/hyperhive/meta/`), the capabilities
+//! the swarm grants (outside the meta repo) and the shared wire-protocol size
 //! limits. Each submodule is re-exported at the crate root, so
 //! `crate::topology::…` etc. keep working unchanged.
 //!
diff --git a/hive-c0re/src/agent_config/resource_limits.rs b/hive-c0re/src/agent_config/resource_limits.rs
index d9e058c9..0eafbfe0 100644
--- a/hive-c0re/src/agent_config/resource_limits.rs
+++ b/hive-c0re/src/agent_config/resource_limits.rs
@@ -3,7 +3,7 @@
 //!
 //! The swarm caps are stored at
 //! `/var/lib/hyperhive/meta/swarm-resource-limits.json` alongside
-//! `topology.json`, `tool-groups.json` and `capabilities.json`.
+//! `topology.json` and `tool-groups.json`.
 //!
 //! Format: a JSON object mapping agent name to an object with optional
 //! `cpu_quota` / `memory_max` strings, passed verbatim to systemd's
diff --git a/hive-c0re/src/agent_config/topology.rs b/hive-c0re/src/agent_config/topology.rs
index d12c297e..cccfe3e5 100644
--- a/hive-c0re/src/agent_config/topology.rs
+++ b/hive-c0re/src/agent_config/topology.rs
@@ -13,7 +13,8 @@
 //! bind-mounted ([`all_agents`]).
 //!
 //! Broader-than-your-own bind-mount grants are **not** stored here: they
-//! hang off the `ManageRootAgent` capability in `capabilities.json`. See
+//! hang off the `ManageRootAgent` capability the swarm grants
+//! (`swarm_permissions`). See
 //! `lifecycle::set_nspawn_flags`.
 
 use std::collections::BTreeSet;
diff --git a/hive-c0re/src/coordinator.rs b/hive-c0re/src/coordinator.rs
index 26815947..50277b1d 100644
--- a/hive-c0re/src/coordinator.rs
+++ b/hive-c0re/src/coordinator.rs
@@ -565,41 +565,6 @@ impl Coordinator {
             .map(|m| m.keys().cloned().collect())
     }
 
-    /// Emit a `CapabilitiesChanged` snapshot event. Called from the
-    /// rebuild-queue worker after a `PermChange` / Capabilities entry
-    /// commits the JSON file, so the P3RM1SS10NS tab updates live.
-    pub fn emit_capabilities_snapshot(self: &Arc<Self>) {
-        use hive_sh4re::permissions::Capability;
-        let caps = Capability::ALL.iter().map(|c| <&str>::from(*c)).collect();
-        let descriptions = Capability::ALL
-            .iter()
-            .map(|c| (<&str>::from(*c), c.description()))
-            .collect();
-        // An empty snapshot would show every agent as having lost its
-        // capabilities; emit nothing and let the HTTP refetch report it.
-        let assignments = match crate::capabilities::read() {
-            Ok(map) => map,
-            Err(e) => {
-                tracing::warn!(error = ?e, "capabilities snapshot skipped");
-                return;
-            }
-        };
-        // Best-effort roster (sync path); on a contended cache miss we
-        // emit explicit keys only — the HTTP refetch fills the rest in.
-        let roster = self.live_container_names_blocking().unwrap_or_default();
-        // Capability default is "no extra caps" — empty default slice.
-        let (agents, effective) =
-            crate::dashboard::permissions::roster_and_effective(roster, &assignments, &[]);
-        self.emit_dashboard_event(DashboardEvent::CapabilitiesChanged {
-            seq: self.next_seq(),
-            caps,
-            descriptions,
-            assignments,
-            agents,
-            effective,
-        });
-    }
-
     /// Emit a `ToolGroupsChanged` snapshot event. Called from the
     /// rebuild-queue worker after a `PermChange` / `ToolGroups` entry
     /// commits the JSON file, so the P3RM1SS10NS tab updates live.
diff --git a/hive-c0re/src/dashboard/mod.rs b/hive-c0re/src/dashboard/mod.rs
index 6f1fec8d..7c9fffd2 100644
--- a/hive-c0re/src/dashboard/mod.rs
+++ b/hive-c0re/src/dashboard/mod.rs
@@ -37,7 +37,7 @@ use crate::lifecycle;
         (name = "lifecycle_ops", description = "agent container lifecycle: rebuild/restart/start/stop/pause/limits"),
         (name = "meta_inputs", description = "bulk flake-input update for the meta flake"),
         (name = "misc_api", description = "mark-all-read, compose, spawn-request, hive stats"),
-        (name = "permissions", description = "tool-group + capability assignment for agents"),
+        (name = "permissions", description = "tool-group assignment for agents"),
         (name = "rebuild_queue", description = "rebuild-queue cancel"),
         (name = "state_files", description = "proxied reads of allow-listed per-agent state files"),
         (name = "state_snapshot", description = "cold-load dashboard snapshot"),
@@ -126,8 +126,6 @@ fn app(coord: Arc<Coordinator>) -> axum::Router {
         .routes(routes!(build_logs::get_build_log_raw))
         .routes(routes!(permissions::get_tool_groups))
         .routes(routes!(permissions::post_tool_groups))
-        .routes(routes!(permissions::get_capabilities))
-        .routes(routes!(permissions::post_capabilities))
         .routes(routes!(permissions::post_permissions))
         .routes(routes!(permissions::get_stale_permissions))
         .routes(routes!(permissions::delete_agent_permissions))
@@ -380,8 +378,6 @@ mod router_build_probe {
             .routes(routes!(build_logs::get_build_log_raw))
             .routes(routes!(permissions::get_tool_groups))
             .routes(routes!(permissions::post_tool_groups))
-            .routes(routes!(permissions::get_capabilities))
-            .routes(routes!(permissions::post_capabilities))
             .routes(routes!(permissions::post_permissions))
             .routes(routes!(permissions::get_stale_permissions))
             .routes(routes!(permissions::delete_agent_permissions))
diff --git a/hive-c0re/src/dashboard/permissions.rs b/hive-c0re/src/dashboard/permissions.rs
index 0ec25b05..7be82410 100644
--- a/hive-c0re/src/dashboard/permissions.rs
+++ b/hive-c0re/src/dashboard/permissions.rs
@@ -1,6 +1,6 @@
-//! Tool-group + capability permission endpoints for the dashboard.
+//! Tool-group permission endpoints for the dashboard.
 //!
-//! Read endpoints return the full set of known groups/capabilities plus
+//! Read endpoints return the full set of known groups plus
 //! descriptions and the per-agent assignment map (the UI never hard-codes
 //! the lists). Write endpoints validate, then enqueue a `PermChange` so the
 //! JSON file write is serialised through the FIFO rebuild worker.
@@ -20,7 +20,7 @@ use super::{AppState, Ident, guard_agent_name, strip_container_prefix};
 #[derive(Serialize, ToSchema)]
 pub(super) struct ToolGroupsSnapshot {
     /// Ordered list of all known tool-group names. Drives the column
-    /// headers in the capabilities table — the UI does not hard-code them.
+    /// headers in the tool-groups table — the UI does not hard-code them.
     groups: Vec<&'static str>,
     /// Short description for each group name. Keys match `groups` entries.
     descriptions: std::collections::BTreeMap<&'static str, &'static str>,
@@ -100,8 +100,8 @@ pub(crate) fn tool_group_default_names() -> Vec<&'static str> {
 /// `agents` is the sorted union of the live roster and any agent that
 /// already has an explicit entry; `effective[agent]` is the explicit
 /// assignment when present, else `default` (the role fallback). Shared by
-/// the capabilities + tool-groups snapshots on both the HTTP and SSE
-/// paths so the two surfaces never drift.
+/// the tool-groups snapshots on both the HTTP and SSE paths so the two
+/// surfaces never drift.
 #[must_use]
 pub(crate) fn roster_and_effective(
     roster: impl IntoIterator<Item = String>,
@@ -184,135 +184,15 @@ pub(super) async fn post_tool_groups(
     Ok((StatusCode::OK, "ok").into_response())
 }
 
-#[derive(Serialize, ToSchema)]
-pub(super) struct CapabilitiesSnapshot {
-    /// Ordered list of all known capability names. Drives the column
-    /// headers in the capabilities table — the UI does not hard-code them.
-    caps: Vec<&'static str>,
-    /// Short description for each capability name. Keys match `caps` entries.
-    descriptions: std::collections::BTreeMap<&'static str, &'static str>,
-    /// Per-agent *explicit* capability grant map. Absent agents have no
-    /// extra caps — the UI uses presence here to badge "(default)".
-    assignments: std::collections::BTreeMap<String, Vec<String>>,
-    /// Every agent the operator can manage (live roster ∪ explicit keys),
-    /// sorted — so default agents always list.
-    agents: Vec<String>,
-    /// Per-agent *effective* caps: explicit entry when present, else the
-    /// default (no caps). Keeps the snapshot shape symmetric with
-    /// tool-groups; the default here is always empty.
-    effective: std::collections::BTreeMap<String, Vec<String>>,
-}
-
-/// Every known capability name + description,
-/// plus the per-agent explicit/effective grant maps.
-#[utoipa::path(
-    get,
-    path = "/api/capabilities",
-    responses(
-        (status = 200, description = "capability catalogue + assignments", body = CapabilitiesSnapshot),
-        (status = 500, description = "capabilities file unreadable"),
-    ),
-    tag = "permissions"
-)]
-pub(super) async fn get_capabilities(
-    State(state): State<AppState>,
-) -> Result<axum::Json<CapabilitiesSnapshot>, ProblemDetails> {
-    use hive_sh4re::permissions::Capability;
-    let caps = Capability::ALL.iter().map(|c| <&str>::from(*c)).collect();
-    let descriptions = Capability::ALL
-        .iter()
-        .map(|c| (<&str>::from(*c), c.description()))
-        .collect();
-    let assignments = crate::capabilities::read().map_err(|e| unreadable(&e))?;
-    let roster = state
-        .coord
-        .containers_snapshot()
-        .await
-        .into_iter()
-        .map(|c| c.name);
-    // Capability default is "no extra caps" — empty default slice.
-    let (agents, effective) = roster_and_effective(roster, &assignments, &[]);
-    Ok(axum::Json(CapabilitiesSnapshot {
-        caps,
-        descriptions,
-        assignments,
-        agents,
-        effective,
-    }))
-}
-
-#[derive(Deserialize, ToSchema)]
-pub(super) struct SetCapabilitiesBody {
-    caps: Vec<String>,
-}
-
-/// Replace `agent`'s explicit capability
-/// grant set (JSON body `{"caps": [...]}`).
-#[utoipa::path(
-    post,
-    path = "/api/capabilities/{agent}",
-    params(("agent" = String, Path, description = "agent name")),
-    request_body = SetCapabilitiesBody,
-    responses(
-        (status = 200, description = "capabilities queued for write", body = String),
-        (status = 400, description = "unknown capability name"),
-        (status = 404, description = "no such agent"),
-    ),
-    tag = "permissions"
-)]
-pub(super) async fn post_capabilities(
-    State(state): State<AppState>,
-    AxumPath(name): AxumPath<String>,
-    axum::Json(body): axum::Json<SetCapabilitiesBody>,
-) -> Result<Response, ProblemDetails> {
-    let logical = strip_container_prefix(&name);
-    if let Some(reject) = guard_agent_name(&state, &logical).await {
-        return Ok(reject);
-    }
-    let known: Vec<&str> = hive_sh4re::permissions::Capability::ALL
-        .iter()
-        .map(|c| <&str>::from(*c))
-        .collect();
-    for cap in &body.caps {
-        if !known.contains(&cap.as_str()) {
-            return Err(ProblemDetails::from_status_code(StatusCode::BAD_REQUEST)
-                .with_detail(format!("unknown capability: {cap}")));
-        }
-    }
-    // Submit a PermChange DAG: the JSON file write commits under
-    // META_LOCK inside the WritePermFile node, so concurrent
-    // batch-apply actions for different agents never race on the
-    // shared capabilities.json.
-    state
-        .coord
-        .job_queue
-        .insert_job(|b| {
-            crate::job_queue::templates::perm_change(
-                b,
-                &logical,
-                crate::job_queue::PermPayload::Capabilities {
-                    caps: body.caps.clone(),
-                },
-            );
-            Vec::new()
-        })
-        .expect("template-declared shapes are acyclic");
-    state.coord.emit_rebuild_queue_snapshot();
-    tracing::info!(agent = %logical, caps = ?body.caps, "operator: set capabilities via dashboard");
-    Ok((StatusCode::OK, "ok").into_response())
-}
-
-/// One agent's slice of a batch permission change. Sparse: an omitted
-/// field leaves that perm-type untouched, an empty array clears it, a
+/// One agent's slice of a batch permission change. An omitted
+/// `tool_groups` leaves the agent untouched, an empty array clears it, a
 /// populated array fully replaces it (same replace semantics as the
-/// per-agent endpoints).
+/// per-agent endpoint).
 #[derive(Deserialize, ToSchema)]
 pub(super) struct PermChangeBody {
     agent: String,
     #[serde(default)]
     tool_groups: Option<Vec<String>>,
-    #[serde(default)]
-    capabilities: Option<Vec<String>>,
 }
 
 #[derive(Deserialize, ToSchema)]
@@ -320,25 +200,19 @@ pub(super) struct BatchPermsBody {
     changes: Vec<PermChangeBody>,
 }
 
-/// A validated, non-empty change staged for enqueue:
-/// `(logical agent, new groups?, new caps?)`.
-type StagedPerm = (String, Option<Vec<String>>, Option<Vec<String>>);
-
 /// Batch permission apply — `POST /api/permissions`.
 ///
-/// The save-all permissions UI sends only the perm-types that actually
-/// changed per agent; each affected agent gets ONE combined
-/// `PermChange`, so the dedup key collapses to `(kind, agent)` and an
-/// agent whose caps AND groups both changed rebuilds once, not twice.
-/// The whole batch is atomic: every change is validated up front and
-/// on any validation error nothing is written or enqueued.
+/// The save-all permissions UI sends only the agents whose groups
+/// changed; each gets one `PermChange`. The whole batch is atomic: every
+/// change is validated up front and on any validation error nothing is
+/// written or enqueued.
 #[utoipa::path(
     post,
     path = "/api/permissions",
     request_body = BatchPermsBody,
     responses(
         (status = 200, description = "batch permission change queued", body = String),
-        (status = 400, description = "unknown tool-group/capability name, or no such agent"),
+        (status = 400, description = "unknown tool-group name, or no such agent"),
         (status = 404, description = "no such agent"),
     ),
     tag = "permissions"
@@ -347,43 +221,25 @@ pub(super) async fn post_permissions(
     State(state): State<AppState>,
     axum::Json(body): axum::Json<BatchPermsBody>,
 ) -> Result<Response, ProblemDetails> {
-    let known_caps: Vec<&str> = hive_sh4re::permissions::Capability::ALL
-        .iter()
-        .map(|c| <&str>::from(*c))
-        .collect();
     // Phase 1 — validate everything before touching any file or the
     // queue, so a bad entry fails the whole POST with zero side effects.
-    // No-op rows (both fields omitted) are skipped, not errors.
-    let mut staged: Vec<StagedPerm> = Vec::new();
+    // No-op rows (field omitted) are skipped, not errors.
+    let mut staged: Vec<(String, Vec<String>)> = Vec::new();
     for change in &body.changes {
         let logical = strip_container_prefix(&change.agent);
         if let Some(reject) = guard_agent_name(&state, &logical).await {
             return Ok(reject);
         }
-        if let Some(groups) = &change.tool_groups
-            && let Err(e) = crate::tool_groups::validate_groups(groups)
-        {
-            return Err(ProblemDetails::from_status_code(StatusCode::BAD_REQUEST)
-                .with_detail(format!("invalid tool-groups for {logical}: {e}")));
-        }
-        if let Some(caps) = &change.capabilities {
-            for cap in caps {
-                if !known_caps.contains(&cap.as_str()) {
-                    return Err(ProblemDetails::from_status_code(StatusCode::BAD_REQUEST)
-                        .with_detail(format!("unknown capability for {logical}: {cap}")));
-                }
+        if let Some(groups) = &change.tool_groups {
+            if let Err(e) = crate::tool_groups::validate_groups(groups) {
+                return Err(ProblemDetails::from_status_code(StatusCode::BAD_REQUEST)
+                    .with_detail(format!("invalid tool-groups for {logical}: {e}")));
             }
-        }
-        if change.tool_groups.is_some() || change.capabilities.is_some() {
-            staged.push((
-                logical,
-                change.tool_groups.clone(),
-                change.capabilities.clone(),
-            ));
+            staged.push((logical, groups.clone()));
         }
     }
-    // Phase 2 — submit one combined PermChange DAG per affected agent.
-    for (logical, groups, caps) in staged {
+    // Phase 2 — submit one PermChange DAG per affected agent.
+    for (logical, groups) in staged {
         state
             .coord
             .job_queue
@@ -391,7 +247,7 @@ pub(super) async fn post_permissions(
                 crate::job_queue::templates::perm_change(
                     b,
                     &logical,
-                    crate::job_queue::PermPayload::Combined { groups, caps },
+                    crate::job_queue::PermPayload::ToolGroups { groups },
                 );
                 Vec::new()
             })
@@ -402,7 +258,7 @@ pub(super) async fn post_permissions(
     Ok((StatusCode::OK, "ok").into_response())
 }
 
-/// Agent names that have explicit capability/tool-group entries but are
+/// Agent names that have explicit tool-group entries but are
 /// not in the live container roster AND not in the kept-state directory
 /// list (i.e. truly gone — renamed or destroyed agents whose JSON entries
 /// persisted). The client uses this to drive the "stale permission entries"
@@ -421,7 +277,7 @@ pub(super) struct StalePermsResponse {
     path = "/api/permissions/stale",
     responses(
         (status = 200, description = "ghost agent names with stale permission entries", body = StalePermsResponse),
-        (status = 500, description = "tool-groups/capabilities file unreadable"),
+        (status = 500, description = "tool-groups file unreadable"),
     ),
     tag = "permissions"
 )]
@@ -447,18 +303,9 @@ pub(super) async fn get_stale_permissions(
             .collect();
     // Known = live roster ∪ kept-state names.
     let known: std::collections::HashSet<&String> = live.iter().chain(kept.iter()).collect();
-    // Explicit entries in either JSON file.
-    let caps = crate::capabilities::read().map_err(|e| unreadable(&e))?;
     let tgs = crate::tool_groups::read().map_err(|e| unreadable(&e))?;
-    let mut ghost_names: Vec<String> = caps
-        .keys()
-        .chain(tgs.keys())
-        .filter(|n| !known.contains(n))
-        .cloned()
-        .collect::<std::collections::HashSet<_>>()
-        .into_iter()
-        .collect();
-    ghost_names.sort();
+    // `tgs` is a `BTreeMap`, so its keys are already sorted.
+    let ghost_names: Vec<String> = tgs.keys().filter(|n| !known.contains(n)).cloned().collect();
     Ok(axum::Json(StalePermsResponse { stale: ghost_names }))
 }
 
@@ -466,15 +313,15 @@ pub(super) async fn get_stale_permissions(
 /// requiring it to exist in the live roster.
 ///
 /// Used by the P3RM1SS10NS tab's "remove" button for agents that have
-/// stale explicit entries in `tool-groups.json` / `capabilities.json`
-/// but are no longer running (e.g. an agent that was renamed or
-/// destroyed while its JSON entries persisted).
+/// stale explicit entries in `tool-groups.json` but are no longer
+/// running (e.g. an agent that was renamed or destroyed while its JSON
+/// entries persisted).
 ///
 /// Bypasses `guard_agent_name`'s live-roster check intentionally —
 /// the whole point is to remove entries for non-roster agents. Only
 /// the format check ([`Ident::parse`]) is applied. No rebuild is
-/// enqueued (the agent doesn't exist to rebuild); the SSE snapshots
-/// update the P3RM1SS10NS tab live.
+/// enqueued (the agent doesn't exist to rebuild); the SSE snapshot
+/// updates the P3RM1SS10NS tab live.
 #[utoipa::path(
     delete,
     path = "/api/permissions/{agent}",
@@ -482,7 +329,7 @@ pub(super) async fn get_stale_permissions(
     responses(
         (status = 200, description = "stale permission entries cleared", body = String),
         (status = 400, description = "bad agent name"),
-        (status = 500, description = "tool-groups/capabilities file write failed"),
+        (status = 500, description = "tool-groups file write failed"),
     ),
     tag = "permissions"
 )]
@@ -496,24 +343,13 @@ pub(super) async fn delete_agent_permissions(
             return (StatusCode::BAD_REQUEST, format!("bad agent name: {reason}")).into_response();
         }
     };
-    // Run both removals regardless so we clean up as much as possible
-    // even on partial I/O errors. Collect errors to surface below.
-    let tg_err = crate::tool_groups::remove_agent(logical.as_str()).err();
-    if let Some(ref e) = tg_err {
-        tracing::warn!(agent = %logical, error = ?e, "failed to remove tool-groups entry");
-    }
-    let cap_err = crate::capabilities::remove_agent(logical.as_str()).err();
-    if let Some(ref e) = cap_err {
-        tracing::warn!(agent = %logical, error = ?e, "failed to remove capabilities entry");
-    }
-    // Emit live snapshots even on partial failure so the UI stays as
-    // accurate as possible — the surviving table gets updated immediately.
+    let removed = crate::tool_groups::remove_agent(logical.as_str());
     state.coord.emit_tool_groups_snapshot();
-    state.coord.emit_capabilities_snapshot();
-    // Surface any I/O error as 500 so the frontend's `!resp.ok` path
+    // Surface an I/O error as 500 so the frontend's `!resp.ok` path
     // fires and the operator sees a meaningful message rather than a
     // silent "success" followed by the row reappearing unchanged.
-    if let Some(e) = tg_err.or(cap_err) {
+    if let Err(e) = removed {
+        tracing::warn!(agent = %logical, error = ?e, "failed to remove tool-groups entry");
         return (
             StatusCode::INTERNAL_SERVER_ERROR,
             format!("failed to remove permission entries for {logical}: {e}"),
diff --git a/hive-c0re/src/dashboard_events.rs b/hive-c0re/src/dashboard_events.rs
index 7cfd5804..0f852c3c 100644
--- a/hive-c0re/src/dashboard_events.rs
+++ b/hive-c0re/src/dashboard_events.rs
@@ -141,24 +141,6 @@ pub enum DashboardEvent {
     /// (transients, container state), and a poll loop would be
     /// slower for the same information.
     RebuildQueueChanged { seq: u64 },
-    /// Full snapshot of capability grants (per-agent `Vec<cap_name>`).
-    /// Emitted from the rebuild-queue worker after a `PermChange`
-    /// `Capabilities` entry commits the JSON file. Lets the P3RM1SS10NS
-    /// tab update live when the worker applies a queued change.
-    CapabilitiesChanged {
-        seq: u64,
-        /// Ordered list of all known capability names (column headers).
-        caps: Vec<&'static str>,
-        /// Short description for each capability name (tooltip).
-        descriptions: std::collections::BTreeMap<&'static str, &'static str>,
-        /// Per-agent *explicit* capability grant map; absent agents have
-        /// no extra caps (the UI badges those "(default)").
-        assignments: std::collections::BTreeMap<String, Vec<String>>,
-        /// Sorted roster the operator can manage (live ∪ explicit keys).
-        agents: Vec<String>,
-        /// Per-agent *effective* caps (explicit-or-default; default empty).
-        effective: std::collections::BTreeMap<String, Vec<String>>,
-    },
     /// Full snapshot of tool-group assignments (per-agent `Vec<group_name>`).
     /// Emitted from the rebuild-queue worker after a `PermChange`
     /// `ToolGroups` entry commits the JSON file. Lets the P3RM1SS10NS
@@ -203,7 +185,6 @@ impl DashboardEvent {
             DashboardEvent::MetaInputsChanged { .. } => "meta_inputs_changed",
             DashboardEvent::MetaUpdateRunning { .. } => "meta_update_running",
             DashboardEvent::RebuildQueueChanged { .. } => "rebuild_queue_changed",
-            DashboardEvent::CapabilitiesChanged { .. } => "capabilities_changed",
             DashboardEvent::ToolGroupsChanged { .. } => "tool_groups_changed",
         }
     }
@@ -279,14 +260,6 @@ mod tests {
                 running: false,
             },
             DashboardEvent::RebuildQueueChanged { seq: 1 },
-            DashboardEvent::CapabilitiesChanged {
-                seq: 1,
-                caps: Vec::new(),
-                descriptions: std::collections::BTreeMap::new(),
-                assignments: std::collections::BTreeMap::new(),
-                agents: Vec::new(),
-                effective: std::collections::BTreeMap::new(),
-            },
             DashboardEvent::ToolGroupsChanged {
                 seq: 1,
                 groups: Vec::new(),
diff --git a/hive-c0re/src/job_queue/exec.rs b/hive-c0re/src/job_queue/exec.rs
index f924aa28..cb08c497 100644
--- a/hive-c0re/src/job_queue/exec.rs
+++ b/hive-c0re/src/job_queue/exec.rs
@@ -810,23 +810,6 @@ async fn run_write_perm_file(
                 .with_context(|| format!("commit tool-groups for {name}"))?;
             coord.emit_tool_groups_snapshot();
         }
-        PermPayload::Capabilities { caps } => {
-            crate::meta::commit_capabilities(name, caps)
-                .await
-                .with_context(|| format!("commit capabilities for {name}"))?;
-            coord.emit_capabilities_snapshot();
-        }
-        PermPayload::Combined { groups, caps } => {
-            crate::meta::commit_perms(name, groups.as_deref(), caps.as_deref())
-                .await
-                .with_context(|| format!("commit perms for {name}"))?;
-            if groups.is_some() {
-                coord.emit_tool_groups_snapshot();
-            }
-            if caps.is_some() {
-                coord.emit_capabilities_snapshot();
-            }
-        }
     }
     Ok(())
 }
diff --git a/hive-c0re/src/job_queue/model.rs b/hive-c0re/src/job_queue/model.rs
index 33e0ce87..97979df5 100644
--- a/hive-c0re/src/job_queue/model.rs
+++ b/hive-c0re/src/job_queue/model.rs
@@ -98,7 +98,7 @@ pub enum NodeKind {
     PauseDrain { agent: String },
     /// `set_nspawn_flags` + `set_resource_limits` + daemon-reload.
     WriteDropin { agent: String },
-    /// Commit `tool-groups.json` / `capabilities.json` per its `payload`.
+    /// Commit `tool-groups.json` per its `payload`.
     WritePermFile { agent: String, payload: PermPayload },
     /// Group root of a rebuild subtree — the **brace**, scoped to one
     /// agent. Why braces exist and what they cost: _Braces_.
diff --git a/hive-c0re/src/job_queue/templates.rs b/hive-c0re/src/job_queue/templates.rs
index b7d4e7e4..3f6ff76d 100644
--- a/hive-c0re/src/job_queue/templates.rs
+++ b/hive-c0re/src/job_queue/templates.rs
@@ -403,8 +403,8 @@ pub fn destroy(builder: &JobBuilder, agent: &str, purge: bool) {
         .after_ok(last);
 }
 
-/// Perm change: commit the JSON file(s), then the rebuild subgraph so
-/// the updated `HIVE_TOOL_GROUPS` / `HIVE_CAPABILITIES` env var takes
+/// Perm change: commit the JSON file, then the rebuild subgraph so
+/// the updated `HIVE_TOOL_GROUPS` env var takes
 /// effect in the container. Group-roots are `WritePermFile` plus the rebuild
 /// subgraph's `MetaSync` / `Prebuild` / `Reconcile`, so the `EmitRebuilt` tail
 /// edges all four.
diff --git a/hive-c0re/src/job_queue/tests.rs b/hive-c0re/src/job_queue/tests.rs
index 8ad97b8d..24165391 100644
--- a/hive-c0re/src/job_queue/tests.rs
+++ b/hive-c0re/src/job_queue/tests.rs
@@ -1522,10 +1522,7 @@ fn perm_change_shape_prefixes_rebuild_chain() {
         templates::perm_change(
             builder,
             "agent-a",
-            PermPayload::Combined {
-                groups: Some(vec![]),
-                caps: None,
-            },
+            PermPayload::ToolGroups { groups: vec![] },
         );
     });
     assert_eq!(
diff --git a/hive-c0re/src/lifecycle/host_config.rs b/hive-c0re/src/lifecycle/host_config.rs
index a0fdd10f..c92564d7 100644
--- a/hive-c0re/src/lifecycle/host_config.rs
+++ b/hive-c0re/src/lifecycle/host_config.rs
@@ -381,22 +381,20 @@ fn bind_own_agent_dirs(agent_name: &str, notes_dir: &Path, binds: &mut Vec<BindM
     }
 }
 
-/// Whether `agent_name` holds `ManageRootAgent` per the capabilities file
-/// at `caps_path`. An unreadable file denies: the error is logged and the
+/// Whether `agent_name` holds `ManageRootAgent` per the swarm permissions
+/// stored at `path`. An unreadable file denies: the error is logged and the
 /// agent starts without the cross-agent mounts, so one corrupt file never
 /// blocks any agent's start, spawn or swap.
-fn holds_manage_root_agent(agent_name: &str, caps_path: &Path) -> bool {
-    crate::capabilities::has_cap(caps_path, agent_name, Capability::ManageRootAgent).unwrap_or_else(
-        |e| {
-            tracing::error!(
-                agent = %agent_name,
-                path = %caps_path.display(),
-                error = ?e,
-                "capabilities unreadable — no ManageRootAgent mounts"
-            );
-            false
-        },
-    )
+fn holds_manage_root_agent(agent_name: &str, path: &Path) -> bool {
+    crate::capabilities::holds(path, agent_name, Capability::ManageRootAgent).unwrap_or_else(|e| {
+        tracing::error!(
+            agent = %agent_name,
+            path = %path.display(),
+            error = ?e,
+            "swarm permissions unreadable — no ManageRootAgent mounts"
+        );
+        false
+    })
 }
 
 /// Env var naming the host directory `swarm-bao-queue-agent.service` lands
@@ -587,7 +585,7 @@ async fn set_nspawn_flags(
     // parent field took with it the unconditional grant every
     // agent used to get over its own direct children — so an agent with
     // no capability now sees its own dirs and nothing else.
-    if holds_manage_root_agent(agent_name, &crate::capabilities::capabilities_path()) {
+    if holds_manage_root_agent(agent_name, &crate::capabilities::path()) {
         // Skipping self is a no-op, not a narrowing: `agent_notes_dir` is
         // `agent_state_dir/state` and `config_bind_source` is shared, so
         // binding the holder as its own virtual child reproduced the two
@@ -818,22 +816,30 @@ mod tests {
     }
 
     #[test]
-    fn a_corrupt_capabilities_file_denies_without_failing() {
+    fn corrupt_swarm_permissions_deny_without_failing() {
         let dir = tempfile::tempdir().expect("tempdir");
-        let path = dir.path().join("capabilities.json");
-        let truncated = "{\n  \"ruth\": [\"manage_root_ag";
+        let path = dir.path().join("swarm-agent-permissions.json");
+        let truncated = "{\n  \"ruth\": {\"capabilities\": [\"manage_root_ag";
         std::fs::write(&path, truncated).expect("seed");
         assert!(!holds_manage_root_agent("ruth", &path));
         assert_eq!(std::fs::read(&path).expect("read"), truncated.as_bytes());
     }
 
+    /// Unset is the default, which gives ruth the recovery mounts and nobody
+    /// else; a grant or an explicit empty list replaces it.
     #[test]
-    fn a_readable_grant_is_honoured() {
+    fn the_swarm_grant_or_the_default_decides_the_recovery_mounts() {
         let dir = tempfile::tempdir().expect("tempdir");
-        let path = dir.path().join("capabilities.json");
-        std::fs::write(&path, r#"{"ruth": ["manage_root_agent"]}"#).expect("seed");
+        let path = dir.path().join("swarm-agent-permissions.json");
         assert!(holds_manage_root_agent("ruth", &path));
         assert!(!holds_manage_root_agent("alice", &path));
+        std::fs::write(
+            &path,
+            r#"{"ruth": {"capabilities": []}, "alice": {"capabilities": ["manage_root_agent"]}}"#,
+        )
+        .expect("seed");
+        assert!(!holds_manage_root_agent("ruth", &path));
+        assert!(holds_manage_root_agent("alice", &path));
     }
 
     const TRUNCATED_LIMITS: &str = "{\n  \"sock\": { \"cpu_quota\": \"50%\", \"mem";
diff --git a/hive-c0re/src/main.rs b/hive-c0re/src/main.rs
index c0ff3bed..bc022efc 100644
--- a/hive-c0re/src/main.rs
+++ b/hive-c0re/src/main.rs
@@ -277,6 +277,9 @@ async fn cmd_serve(
     if let Err(e) = migrate::run(&coord).await {
         tracing::warn!(error = ?e, "startup migration failed");
     }
+    // Before any agent starts, so the bind mounts it gets from the stored
+    // permissions are the meta repo's grants until the swarm answers.
+    crate::workers::permissions_import::seed_local();
     // Auto-create the root agent container if it isn't there yet. Block
     // on this — without root the system has no manager harness.
     // Failures are logged but allowed: a broken auto-spawn shouldn't
diff --git a/hive-c0re/src/meta.rs b/hive-c0re/src/meta.rs
index ce916fa8..c57559a6 100644
--- a/hive-c0re/src/meta.rs
+++ b/hive-c0re/src/meta.rs
@@ -47,8 +47,8 @@ pub struct AgentSpec {
 
 /// Stage every generated meta JSON file that exists: topology.json is
 /// regenerated by `reconcile` whenever the agent set changed;
-/// tool-groups/capabilities are created lazily on
-/// first write (`set_groups`/`set_caps`) — absent means
+/// tool-groups is created lazily on
+/// first write (`set_groups`) — absent means
 /// every agent is on defaults, no file needed. Without staging, an
 /// existing-but-untracked file shows up as untracked in the meta repo,
 /// which can confuse nix's dirty-tree fetch. `git add` is a no-op when
@@ -57,10 +57,6 @@ async fn stage_generated_meta_files(dir: &std::path::Path) -> Result<()> {
     for (path, name) in [
         (crate::topology::topology_path(), "topology.json"),
         (crate::tool_groups::tool_groups_path(), "tool-groups.json"),
-        (
-            crate::capabilities::capabilities_path(),
-            "capabilities.json",
-        ),
     ] {
         if path.exists() {
             git(dir, &["add", name]).await?;
@@ -197,7 +193,6 @@ pub async fn sync_agents(hive: &HiveEnv, agents: &[AgentSpec]) -> Result<()> {
                 // Only ever a REMOVAL now; see `is_embedded_ca_name`.
                 f if f.starts_with("peer-ca-") && has_pem_ext(f) => Some("peer-ca"),
                 "topology.json" => Some("topology"),
-                "capabilities.json" => Some("capabilities"),
                 "tool-groups.json" => Some("tool-groups"),
                 _ => None,
             })
@@ -326,68 +321,28 @@ pub async fn commit_tool_groups(agent: &str, groups: &[String]) -> Result<()> {
     Ok(())
 }
 
-/// Write the capabilities file for `agent` and commit it atomically
-/// under `META_LOCK`. Same rationale as `commit_tool_groups`.
-pub async fn commit_capabilities(agent: &str, caps: &[String]) -> Result<()> {
-    let _guard = META_LOCK.lock().await;
-    crate::capabilities::set_caps(agent, caps)
-        .map_err(|e| anyhow::anyhow!("set capabilities for {agent}: {e}"))?;
-    let dir = crate::paths::meta_root();
-    if crate::capabilities::capabilities_path().exists() {
-        git(&dir, &["add", "capabilities.json"]).await?;
-    }
-    if paths_dirty(&dir, &["capabilities.json"]).await? {
-        git_commit_paths(
-            &dir,
-            &format!("set capabilities for {agent}"),
-            &["capabilities.json"],
-        )
-        .await?;
-    }
-    Ok(())
-}
-
-/// Write both perm files for `agent` (whichever are `Some`) and commit
-/// them in a SINGLE git commit under `META_LOCK` — the batch
-/// `POST /api/permissions` path. A `None` field leaves that file
-/// untouched. One commit + (caller does) one rebuild means changing an
-/// agent's caps and tool-groups together no longer triggers two
-/// rebuilds. Mirrors the staging discipline of `commit_tool_groups` /
-/// `commit_capabilities`.
+/// Delete `capabilities.json` from the meta repo and commit the removal under
+/// `META_LOCK`. Its grants live in the swarm's `agent-permissions` bucket
+/// once `workers::permissions_import` has sent them.
 ///
 /// # Errors
 ///
-/// Returns an error if writing either JSON file fails, a capability name
-/// is invalid (`set_caps`), or a git stage/commit step fails.
-pub async fn commit_perms(
-    agent: &str,
-    groups: Option<&[String]>,
-    caps: Option<&[String]>,
-) -> Result<()> {
+/// Removing the file or a git stage/commit step fails.
+pub async fn remove_capabilities_file() -> Result<()> {
+    const FILE: &str = crate::workers::permissions_import::META_FILE;
     let _guard = META_LOCK.lock().await;
     let dir = crate::paths::meta_root();
-    let mut parts: Vec<&str> = Vec::new();
-    if let Some(groups) = groups {
-        crate::tool_groups::set_groups(agent, groups)?;
-        if crate::tool_groups::tool_groups_path().exists() {
-            git(&dir, &["add", "tool-groups.json"]).await?;
-        }
-        parts.push("tool-groups");
+    match std::fs::remove_file(dir.join(FILE)) {
+        Ok(()) => {}
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
+        Err(e) => return Err(e).with_context(|| format!("removing {FILE}")),
     }
-    if let Some(caps) = caps {
-        crate::capabilities::set_caps(agent, caps)
-            .map_err(|e| anyhow::anyhow!("set capabilities for {agent}: {e}"))?;
-        if crate::capabilities::capabilities_path().exists() {
-            git(&dir, &["add", "capabilities.json"]).await?;
-        }
-        parts.push("capabilities");
-    }
-    let paths = ["tool-groups.json", "capabilities.json"];
-    if paths_dirty(&dir, &paths).await? {
+    git(&dir, &["rm", "--cached", "--ignore-unmatch", "-q", FILE]).await?;
+    if paths_dirty(&dir, &[FILE]).await? {
         git_commit_paths(
             &dir,
-            &format!("set {} for {agent}", parts.join(" + ")),
-            &paths,
+            "capabilities moved to the swarm's agent-permissions",
+            &[FILE],
         )
         .await?;
     }
@@ -905,7 +860,7 @@ where
     let pronouns_escaped = operator_pronouns.replace('\\', "\\\\").replace('"', "\\\"");
     let _ = writeln!(
         out,
-        "      dashboardPort = {dashboard_port};\n      operatorPronouns = \"{pronouns_escaped}\";\n      mkAgent = {{ name, isManager, port, toolGroups ? null, capabilities ? null }}:"
+        "      dashboardPort = {dashboard_port};\n      operatorPronouns = \"{pronouns_escaped}\";\n      mkAgent = {{ name, isManager, port, toolGroups ? null }}:"
     );
     out.push_str(
         r#"        let
@@ -915,7 +870,6 @@ where
           input = inputs."agent-${name}";
           service = "hive-agent";
           toolGroupsEnv = if toolGroups == null then {} else { HIVE_TOOL_GROUPS = toolGroups; };
-          capabilitiesEnv = if capabilities == null then {} else { HIVE_CAPABILITIES = capabilities; };
         in
         base.extendModules {
           modules = [
@@ -1092,7 +1046,7 @@ where
     }
     out.push_str(
         r"              };
-              systemd.services.${service}.environment = toolGroupsEnv // capabilitiesEnv // {
+              systemd.services.${service}.environment = toolGroupsEnv // {
                 HIVE_PORT = toString port;
                 HIVE_LABEL = name;
                 HIVE_DASHBOARD_PORT = toString dashboardPort;
@@ -1140,7 +1094,6 @@ where
     // renders `toolGroups = null` and gets `AGENT_DEFAULT`, which fails
     // open for any agent whose explicit entry is narrower.
     let tool_groups_map = crate::tool_groups::read()?;
-    let capabilities_map = crate::capabilities::read()?;
     for spec in agents {
         // Emit `toolGroups = "group1,group2"` when the operator has
         // explicitly configured groups for this agent. Absent entry = null
@@ -1153,28 +1106,14 @@ where
             let joined = groups.join(",");
             format!("\"{joined}\"")
         };
-        // Emit `capabilities = "cap1,cap2"` when the operator has
-        // granted capabilities to this agent. Absent entry = null = no
-        // capability env var injected, capability-gated tools hidden.
-        let caps = capabilities_map
-            .get(&spec.name)
-            .cloned()
-            .unwrap_or_default();
-        let capabilities_attr = if caps.is_empty() {
-            "null".to_owned()
-        } else {
-            let joined = caps.join(",");
-            format!("\"{joined}\"")
-        };
         let _ = writeln!(
             out,
-            "        {} = mkAgent {{ name = \"{}\"; isManager = {}; port = {}; toolGroups = {}; capabilities = {}; }};",
+            "        {} = mkAgent {{ name = \"{}\"; isManager = {}; port = {}; toolGroups = {}; }};",
             spec.name,
             spec.name,
             if spec.is_manager { "true" } else { "false" },
             spec.port,
             tool_groups_attr,
-            capabilities_attr,
         );
     }
     out.push_str("      };\n    };\n}\n");
diff --git a/hive-c0re/src/swarm_status.rs b/hive-c0re/src/swarm_status.rs
index c4cfc54b..c24a5ba5 100644
--- a/hive-c0re/src/swarm_status.rs
+++ b/hive-c0re/src/swarm_status.rs
@@ -98,7 +98,7 @@ pub fn spawn(
         // reason, and is a third consumer rather than a second connection.
         tokio::spawn(crate::workers::wanted::watch_declarations(
             client.clone(),
-            coord,
+            std::sync::Arc::clone(&coord),
             hive.clone(),
             shutdown.clone(),
         ));
@@ -109,6 +109,14 @@ pub fn spawn(
             shutdown.clone(),
         ));
 
+        // The capabilities the swarm grants, and the import of the meta repo's.
+        tokio::spawn(crate::workers::swarm_permissions::follow(
+            client.clone(),
+            std::sync::Arc::clone(&coord),
+            hive.clone(),
+            shutdown.clone(),
+        ));
+
         let mut health = SweepHealth::new("swarm_status_publish", "warn", FAILURES_BEFORE_BANNER);
         loop {
             match publish(&client, &hive).await {
diff --git a/hive-c0re/src/workers/auto_update.rs b/hive-c0re/src/workers/auto_update.rs
index ac31b48f..029032f7 100644
--- a/hive-c0re/src/workers/auto_update.rs
+++ b/hive-c0re/src/workers/auto_update.rs
@@ -124,10 +124,6 @@ pub async fn ensure_root_agent(coord: &Arc<Coordinator>) -> Result<()> {
         );
         return Ok(());
     }
-    // Before the create/start branch, not inside it: on a hive whose root
-    // container already exists this is the only run that can still seed the
-    // grant, and it has to land before her next rebuild bakes the binds.
-    seed_manager_capabilities();
     let list_result = lifecycle::list().await;
     if let Err(e) = &list_result {
         tracing::warn!(
@@ -257,68 +253,6 @@ fn seed_manager_tool_groups() {
     }
 }
 
-/// The capability the root agent is seeded with, as the `snake_case` string
-/// `capabilities.json` stores and `capabilities::has_cap` compares against.
-/// Taken from the enum rather than written out, so the seed cannot drift into
-/// a name `capabilities::prune_unknown` would silently drop.
-fn manager_seed_caps() -> Vec<String> {
-    vec![<&str>::from(hive_sh4re::permissions::Capability::ManageRootAgent).to_owned()]
-}
-
-/// Whether the root agent's default capability grant should be written.
-///
-/// `store_written` is "`capabilities.json` exists" — see
-/// [`seed_manager_capabilities`] for why that, and not the absence of an
-/// entry for the manager, is the condition.
-fn should_seed_manager_caps(store_written: bool) -> bool {
-    !store_written
-}
-
-/// Give ruth the `ManageRootAgent` capability on the path that deploys her.
-///
-/// This is the capability-store half of the grant `roles.json` used to make:
-/// `topology::reconcile_roles` seeded `can_manage_top_level_agents` onto
-/// `MANAGER_NAME` on every meta sync, and that role is what put the other
-/// agents' state/config dirs, `/applied` and `/meta` into her nspawn binds.
-/// Collapsing the role into the capability means that without a seed here she
-/// loses those recovery mounts at her next container rebuild — silently, and
-/// only then, because nspawn bakes bind flags at container start.
-///
-/// **Seeded once, not re-ensured on every boot.** The role kept an empty-list
-/// tombstone so an explicit revoke stuck; the capability store *deletes* an
-/// entry that has been emptied rather than tombstoning it, so "the manager
-/// has no entry" cannot tell a fresh hive apart from a deliberate revoke.
-/// File existence can: every grant and revoke goes through
-/// `meta::commit_capabilities` → `capabilities::set_caps`, which
-/// writes the file even when the result is an empty `{}`. So while the file
-/// is absent nobody has ever had a say, and once it exists this is inert
-/// forever — including on the destroy+recreate path, matching
-/// [`seed_manager_tool_groups`]'s refusal to reset an operator's choice.
-///
-/// Written with plain `set_caps` rather than `meta::commit_capabilities`: the
-/// role's own seed wrote the file and left it for `stage_generated_meta_files`
-/// to `git add` into the next deploy commit, and taking `META_LOCK` on the
-/// boot path would be a new ordering constraint for no gain.
-fn seed_manager_capabilities() {
-    if !should_seed_manager_caps(crate::capabilities::capabilities_path().exists()) {
-        tracing::debug!(
-            "capabilities.json already written — leaving the manager's capabilities as-is"
-        );
-        return;
-    }
-    let caps = manager_seed_caps();
-    match crate::capabilities::set_caps(MANAGER_NAME, &caps) {
-        Ok(()) => tracing::info!(
-            caps = ?caps,
-            "seeded ruth's default capabilities (recovery bind mounts for every agent)"
-        ),
-        Err(e) => tracing::warn!(
-            error = ?e,
-            "failed to seed ruth's capabilities — she will come up without the recovery mounts"
-        ),
-    }
-}
-
 /// Boot reconcile (see the module doc): classify every agent by rev
 /// freshness + persisted `wanted` intent, submit one `Boot` DAG
 /// (hyperhive lock bump growing an in-DAG rebuild subgraph per stale
@@ -572,10 +506,7 @@ fn submit_boot_tree(
 
 #[cfg(test)]
 mod tests {
-    use super::{
-        BootAction, MANAGER_NAME, RootAgentPlan, boot_action, manager_seed_caps, plan_root_agent,
-        should_seed_manager_caps,
-    };
+    use super::{BootAction, MANAGER_NAME, RootAgentPlan, boot_action, plan_root_agent};
     use crate::lifecycle::AGENT_PREFIX;
     use crate::power::Wanted;
 
@@ -604,45 +535,6 @@ mod tests {
         assert_eq!(plan_root_agent(result), RootAgentPlan::Present);
     }
 
-    // -----------------------------------------------------------------------
-    // Root-agent capability seed. `capabilities_path()` resolves under
-    // `paths::meta_root()`, which is hardcoded to `/var/lib/hyperhive` with no
-    // test override, so — same as `capabilities.rs`'s own tests — these pin
-    // the pure decision (`should_seed_manager_caps`, `manager_seed_caps`)
-    // rather than round-tripping the real file. Together they are the
-    // replacement for the deleted `reconcile_roles_in_seeds_root_when_absent`
-    // and `reconcile_roles_in_does_not_reseed_after_explicit_revoke`.
-
-    #[test]
-    fn the_manager_is_seeded_while_the_capability_store_has_never_been_written() {
-        assert!(should_seed_manager_caps(false));
-        assert_eq!(manager_seed_caps(), vec!["manage_root_agent".to_owned()]);
-    }
-
-    /// The revoke-sticks half, and the reason the condition is the file rather
-    /// than the manager's entry: the store deletes an emptied entry instead of
-    /// keeping a tombstone, but a revoke still writes the file — so once it
-    /// exists the seed must never fire again.
-    #[test]
-    fn a_written_capability_store_is_never_reseeded() {
-        assert!(!should_seed_manager_caps(true));
-    }
-
-    /// A seed the store would drop as unrecognised grants nothing while
-    /// logging success, so pin that what we write is a name `prune_unknown`
-    /// keeps — i.e. one of `Capability::ALL`'s `snake_case` spellings.
-    #[test]
-    fn the_seeded_name_is_a_capability_the_store_recognises() {
-        for cap in manager_seed_caps() {
-            assert!(
-                hive_sh4re::permissions::Capability::ALL
-                    .iter()
-                    .any(|known| <&str>::from(*known) == cap),
-                "{cap} is not a known capability"
-            );
-        }
-    }
-
     // The regression the `Unreadable` variant exists for: the arm used to
     // substitute `from_running(running)`, which is `Noop` against BOTH
     // observations — so no combination of inputs could reach a reconcile.
diff --git a/hive-c0re/src/workers/mod.rs b/hive-c0re/src/workers/mod.rs
index 6f2ffa4c..118f923b 100644
--- a/hive-c0re/src/workers/mod.rs
+++ b/hive-c0re/src/workers/mod.rs
@@ -1,7 +1,7 @@
 //! Background tasks and periodic sweeps: crash/login watcher, boot-time
 //! auto-update reconcile, the agent-sockets.json writer loop, the MCP socket listener
 //! reconcile loop, knowledge-repo sync, the swarm wanted-state pull, and the
-//! swarm resource-cap read.
+//! swarm resource-cap and permissions reads.
 //! Each submodule is re-exported
 //! at the crate root, so `crate::crash_watch::…` etc. keep working unchanged.
 
@@ -10,5 +10,7 @@ pub mod auto_update;
 pub mod crash_watch;
 pub mod knowledge;
 pub mod mcp_sockets;
+pub mod permissions_import;
 pub mod swarm_limits;
+pub mod swarm_permissions;
 pub mod wanted;
diff --git a/hive-c0re/src/workers/permissions_import.rs b/hive-c0re/src/workers/permissions_import.rs
new file mode 100644
index 00000000..a0463e3b
--- /dev/null
+++ b/hive-c0re/src/workers/permissions_import.rs
@@ -0,0 +1,197 @@
+//! The one-shot move of the capabilities this hive's meta repo held
+//! (`meta/capabilities.json`) into the swarm's `agent-permissions` bucket.
+//!
+//! At boot, before any agent starts, [`seed_local`] copies the file's grants
+//! into [`crate::capabilities`] when nothing is stored there yet, so agents
+//! keep their mounts until the swarm answers. [`send_pending`] then sends the
+//! grants to swarm-controller, which writes each agent's key only where there
+//! is none, and on its answer deletes the file from the meta repo. Nothing
+//! reads the file after that.
+
+use std::collections::BTreeMap;
+use std::path::{Path, PathBuf};
+
+use anyhow::{Context, Result};
+use hive_sh4re::permissions::Capability;
+use swarm_queue_client::agent_permissions::{
+    self, AgentPermissions, IMPORT_REPLY_DEADLINE, Import, ImportReply,
+};
+
+/// The meta repo file the grants are moved out of.
+pub const META_FILE: &str = "capabilities.json";
+
+fn meta_path() -> PathBuf {
+    crate::paths::meta_root().join(META_FILE)
+}
+
+/// The grants the meta file holds, by agent, as the swarm stores them. A name
+/// that is not a [`Capability`] is dropped, with a warning.
+///
+/// `ruth` with no entry in a file that exists had her grant revoked: the file
+/// was created holding it, and a revoke removed the entry. She is sent an
+/// empty list, since unset would give her the default back.
+fn grants(file: BTreeMap<String, Vec<String>>) -> BTreeMap<String, Vec<Capability>> {
+    let mut out: BTreeMap<String, Vec<Capability>> = file
+        .into_iter()
+        .map(|(agent, names)| {
+            let caps = names
+                .iter()
+                .filter_map(|name| {
+                    let cap = Capability::ALL
+                        .iter()
+                        .copied()
+                        .find(|c| name.eq_ignore_ascii_case(<&str>::from(*c)));
+                    if cap.is_none() {
+                        tracing::warn!(%agent, capability = %name, "import: dropping unrecognised capability");
+                    }
+                    cap
+                })
+                .collect();
+            (agent, caps)
+        })
+        .collect();
+    out.entry(crate::lifecycle::MANAGER_NAME.to_owned())
+        .or_default();
+    out
+}
+
+fn read_grants(path: &Path) -> Result<Option<BTreeMap<String, Vec<Capability>>>> {
+    if !path.exists() {
+        return Ok(None);
+    }
+    let file: BTreeMap<String, Vec<String>> = serde_json::from_slice(
+        &std::fs::read(path).with_context(|| format!("reading {}", path.display()))?,
+    )
+    .with_context(|| format!("parsing {}", path.display()))?;
+    Ok(Some(grants(file)))
+}
+
+/// Store the meta file's grants as the last-read swarm permissions, when the
+/// file is still there and nothing has been stored yet.
+pub fn seed_local() {
+    if let Err(e) = seed_local_at(&meta_path(), &crate::capabilities::path()) {
+        tracing::warn!(error = ?e, "import: seeding the stored permissions from the meta repo failed");
+    }
+}
+
+fn seed_local_at(meta: &Path, stored: &Path) -> Result<()> {
+    if stored.exists() {
+        return Ok(());
+    }
+    let Some(grants) = read_grants(meta)? else {
+        return Ok(());
+    };
+    let records = grants
+        .into_iter()
+        .map(|(agent, caps)| {
+            (
+                agent,
+                AgentPermissions {
+                    capabilities: Some(caps),
+                },
+            )
+        })
+        .collect();
+    crate::capabilities::write_at(stored, &records).context("storing the seeded permissions")?;
+    tracing::info!("import: seeded the stored permissions from the meta repo's capabilities");
+    Ok(())
+}
+
+/// Send the meta file's grants to swarm-controller while the file is there,
+/// and delete it once the controller has answered that it handled them.
+///
+/// # Errors
+///
+/// The file is still there: the import failed, or deleting the file did. The
+/// next pass sends again; the controller keeps a key it already has.
+pub async fn send_pending(client: &async_nats::Client, hive: &str) -> Result<()> {
+    send_at(client, hive, &meta_path())
+        .await
+        .context("importing the meta repo's capabilities")
+}
+
+async fn send_at(client: &async_nats::Client, hive: &str, path: &Path) -> Result<()> {
+    let Some(capabilities) = read_grants(path)? else {
+        return Ok(());
+    };
+    swarm_queue_client::ensure_connected(client)?;
+    let payload = serde_json::to_vec(&Import { capabilities })?;
+    let reply = tokio::time::timeout(
+        IMPORT_REPLY_DEADLINE,
+        client.request(agent_permissions::import_subject(hive), payload.into()),
+    )
+    .await
+    .context("the controller did not answer the import")?
+    .context("sending the import")?;
+    match serde_json::from_slice(&reply.payload).context("decoding the controller's answer")? {
+        ImportReply::Imported {
+            written,
+            kept,
+            skipped,
+        } => {
+            tracing::info!(
+                ?written,
+                ?kept,
+                ?skipped,
+                "import: the controller took the meta repo's capabilities"
+            );
+            crate::meta::remove_capabilities_file()
+                .await
+                .context("deleting the imported capabilities file")
+        }
+        ImportReply::Failed { error } => {
+            anyhow::bail!("the controller refused the import: {error}")
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+
+    use super::{grants, seed_local_at};
+    use hive_sh4re::permissions::Capability;
+
+    fn file(json: &str) -> BTreeMap<String, Vec<String>> {
+        serde_json::from_str(json).expect("decodes")
+    }
+
+    #[test]
+    fn grants_keep_known_names_and_a_revoked_ruth_stays_revoked() {
+        let out = grants(file(
+            r#"{"atlas": ["manage_root_agent", "fly_to_the_moon"], "iris": []}"#,
+        ));
+        assert_eq!(out["atlas"], vec![Capability::ManageRootAgent]);
+        assert_eq!(out["iris"], Vec::<Capability>::new());
+        assert_eq!(out["ruth"], Vec::<Capability>::new());
+
+        let out = grants(file(r#"{"ruth": ["manage_root_agent"]}"#));
+        assert_eq!(out["ruth"], vec![Capability::ManageRootAgent]);
+    }
+
+    /// The seed writes once, from the file, and never over a stored read.
+    #[test]
+    fn the_seed_copies_the_file_only_when_nothing_is_stored() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let meta = dir.path().join("capabilities.json");
+        let stored = dir.path().join("stored.json");
+
+        seed_local_at(&meta, &stored).expect("no file is no seed");
+        assert!(!stored.exists());
+
+        std::fs::write(&meta, r#"{"atlas": ["manage_root_agent"]}"#).expect("seed");
+        seed_local_at(&meta, &stored).expect("seeded");
+        assert!(
+            crate::capabilities::holds(&stored, "atlas", Capability::ManageRootAgent)
+                .expect("read")
+        );
+        assert!(
+            !crate::capabilities::holds(&stored, "ruth", Capability::ManageRootAgent)
+                .expect("read")
+        );
+
+        std::fs::write(&stored, "{}").expect("a later read");
+        seed_local_at(&meta, &stored).expect("left alone");
+        assert_eq!(std::fs::read_to_string(&stored).expect("read"), "{}");
+    }
+}
diff --git a/hive-c0re/src/workers/swarm_permissions.rs b/hive-c0re/src/workers/swarm_permissions.rs
new file mode 100644
index 00000000..7467d61e
--- /dev/null
+++ b/hive-c0re/src/workers/swarm_permissions.rs
@@ -0,0 +1,295 @@
+//! Reading the capabilities the swarm grants this hive's agents.
+//!
+//! swarm-controller keeps one key per agent in the `agent-permissions` bucket
+//! ([`swarm_queue_client::agent_permissions`]). Every [`READ_INTERVAL`] this
+//! reads the key of each agent this hive has and stores what it finds with
+//! [`crate::capabilities`], where the container's bind mounts are worked out.
+//! Those are fixed when the container starts, so an agent whose capabilities
+//! changed is restarted, after its current turn.
+//!
+//! A failed read, or a bucket that does not answer, leaves the stored
+//! permissions as they are and restarts nothing: dropping them would put an
+//! agent back on its default, which can be wider than what the swarm set.
+
+use std::path::Path;
+use std::sync::Arc;
+use std::time::Duration;
+
+use anyhow::{Context, Result};
+use swarm_queue_client::agent_permissions::{self, AgentPermissions, Read};
+
+use crate::capabilities::{self, Stored};
+use crate::coordinator::Coordinator;
+use crate::stats::sweep_health::{self, SweepHealth};
+
+/// How often the permissions are read.
+pub const READ_INTERVAL: Duration = Duration::from_secs(30);
+
+/// At [`READ_INTERVAL`] this is ~90 seconds of failing reads.
+const FAILURES_BEFORE_BANNER: u32 = 3;
+
+/// Send the meta repo's grants once, then read the permissions now and every
+/// [`READ_INTERVAL`] until shutdown.
+pub async fn follow(
+    client: async_nats::Client,
+    coord: Arc<Coordinator>,
+    hive: String,
+    mut shutdown: tokio::sync::watch::Receiver<bool>,
+) {
+    let mut health = SweepHealth::new("swarm_permissions_read", "warn", FAILURES_BEFORE_BANNER);
+    loop {
+        // While the meta repo's grants are not in the swarm yet, an unset key
+        // is not the swarm's answer, so nothing is read until they are.
+        let pass = match super::permissions_import::send_pending(&client, &hive).await {
+            Ok(()) => sync(&client).await,
+            Err(e) => Err(e),
+        };
+        match pass {
+            Ok(changed) => {
+                health.record_ok();
+                restart(&coord, changed).await;
+            }
+            Err(e) => {
+                tracing::warn!(error = ?e, "swarm permissions: the pass failed");
+                let err = format!("{e:#}");
+                health.record_err(|ctx| {
+                    let age = ctx.since_last_ok.map_or_else(
+                        || "no success this session".to_owned(),
+                        |d| format!("last ok {} ago", sweep_health::fmt_age(d)),
+                    );
+                    format!(
+                        "syncing the swarm's agent permissions is failing ({} consecutive, \
+                         {age}) — agents keep the permissions last read: {err}",
+                        ctx.consecutive
+                    )
+                });
+            }
+        }
+        tokio::select! {
+            () = tokio::time::sleep(READ_INTERVAL) => {}
+            _ = shutdown.changed() => {
+                tracing::info!("swarm permissions: shutdown signal received");
+                return;
+            }
+        }
+    }
+}
+
+/// Read every agent's key and store the result. Returns the agents whose
+/// capabilities changed.
+async fn sync(client: &async_nats::Client) -> Result<Vec<String>> {
+    // An unconnected client does not fail a JetStream request, it hangs on it.
+    swarm_queue_client::ensure_connected(client)?;
+    let agents: Vec<String> = crate::lifecycle::agents_for_meta_listing()
+        .await
+        .context("enumerating this hive's agents")?
+        .into_iter()
+        .map(|spec| spec.name)
+        .collect();
+    sync_with(&agents, &capabilities::path(), |agent| async move {
+        agent_permissions::read(client, &agent)
+            .await
+            .map_err(|e| anyhow::anyhow!(e))
+            .with_context(|| format!("reading the swarm's permissions for {agent}"))
+    })
+    .await
+}
+
+/// [`sync`] with the read and the file given. Nothing is written unless every
+/// read succeeds.
+async fn sync_with<F, Fut>(agents: &[String], path: &Path, read: F) -> Result<Vec<String>>
+where
+    F: Fn(String) -> Fut,
+    Fut: std::future::Future<Output = Result<Read>>,
+{
+    let mut read_now = Stored::new();
+    for agent in agents {
+        match read(agent.clone()).await? {
+            Read::Current(record) => {
+                if record != AgentPermissions::default() {
+                    read_now.insert(agent.clone(), record);
+                }
+            }
+            Read::NoBucket => {
+                tracing::debug!("swarm permissions: the agent-permissions bucket does not answer");
+                return Ok(Vec::new());
+            }
+        }
+    }
+    let stored = capabilities::read_at(path).context("reading the stored swarm permissions")?;
+    if read_now == stored {
+        return Ok(Vec::new());
+    }
+    capabilities::write_at(path, &read_now).context("storing the swarm permissions")?;
+    tracing::info!(permissions = ?read_now, "swarm permissions: stored the swarm's permissions");
+    Ok(agents
+        .iter()
+        .filter(|a| {
+            capabilities::effective(a, stored.get(*a))
+                != capabilities::effective(a, read_now.get(*a))
+        })
+        .cloned()
+        .collect())
+}
+
+/// Restart each running agent in `changed` once its current turn is done, so
+/// its container starts with the bind mounts its new capabilities give. A
+/// stopped agent gets them at its next start.
+async fn restart(coord: &Arc<Coordinator>, changed: Vec<String>) {
+    let mut running = Vec::new();
+    for agent in changed {
+        if crate::lifecycle::is_running(&agent).await {
+            running.push(agent);
+        }
+    }
+    if running.is_empty() {
+        return;
+    }
+    tracing::info!(agents = ?running, "swarm permissions: capabilities changed, restarting");
+    if let Err(e) = crate::job_queue::power::restart_many(coord, &running, true).await {
+        tracing::error!(agents = ?running, error = ?e, "swarm permissions: restart insert failed");
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::path::Path;
+
+    use super::{Read, sync_with};
+    use swarm_queue_client::agent_permissions::AgentPermissions;
+
+    fn record(json: &str) -> AgentPermissions {
+        serde_json::from_str(json).expect("decodes")
+    }
+
+    fn agents() -> Vec<String> {
+        vec!["atlas".to_owned(), "ruth".to_owned()]
+    }
+
+    /// Sync `agents()` against `path` with each agent read as `answer` gives.
+    async fn sync(
+        path: &Path,
+        answer: impl Fn(&str) -> anyhow::Result<Read>,
+    ) -> anyhow::Result<Vec<String>> {
+        sync_with(&agents(), path, |agent| {
+            let out = answer(&agent);
+            async move { out }
+        })
+        .await
+    }
+
+    /// No key anywhere is everyone on the default: nothing stored, and no
+    /// restart, since that is what the agents already run with.
+    #[tokio::test]
+    async fn missing_keys_are_the_default_and_restart_nothing() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let path = dir.path().join("p.json");
+        let changed = sync(&path, |_| Ok(Read::Current(AgentPermissions::default())))
+            .await
+            .expect("read");
+        assert!(changed.is_empty());
+        assert!(
+            crate::capabilities::read_at(&path)
+                .expect("read")
+                .is_empty()
+        );
+        assert!(
+            crate::capabilities::holds(
+                &path,
+                "ruth",
+                hive_sh4re::permissions::Capability::ManageRootAgent
+            )
+            .expect("read")
+        );
+    }
+
+    /// Revoking ruth's default, and granting atlas, restarts both; clearing
+    /// them again restarts both back.
+    #[tokio::test]
+    async fn a_changed_capability_asks_for_a_restart() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let path = dir.path().join("p.json");
+        let changed = sync(&path, |agent| {
+            Ok(Read::Current(record(if agent == "ruth" {
+                r#"{"capabilities":[]}"#
+            } else {
+                r#"{"capabilities":["manage_root_agent"]}"#
+            })))
+        })
+        .await
+        .expect("read");
+        assert_eq!(changed, agents());
+
+        // The same answer again changes nothing.
+        let unchanged = sync(&path, |agent| {
+            Ok(Read::Current(record(if agent == "ruth" {
+                r#"{"capabilities":[]}"#
+            } else {
+                r#"{"capabilities":["manage_root_agent"]}"#
+            })))
+        })
+        .await
+        .expect("read");
+        assert!(unchanged.is_empty());
+
+        let cleared = sync(&path, |_| Ok(Read::Current(AgentPermissions::default())))
+            .await
+            .expect("read");
+        assert_eq!(cleared, agents());
+    }
+
+    /// Setting ruth's default explicitly is a different record but the same
+    /// mounts, so it is stored and restarts nothing.
+    #[tokio::test]
+    async fn an_explicit_default_is_stored_without_a_restart() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let path = dir.path().join("p.json");
+        let changed = sync(&path, |agent| {
+            Ok(Read::Current(if agent == "ruth" {
+                record(r#"{"capabilities":["manage_root_agent"]}"#)
+            } else {
+                AgentPermissions::default()
+            }))
+        })
+        .await
+        .expect("read");
+        assert!(changed.is_empty());
+        assert!(
+            crate::capabilities::read_at(&path)
+                .expect("read")
+                .contains_key("ruth")
+        );
+    }
+
+    /// One failed read, or a bucket that does not answer, keeps what was
+    /// stored and restarts nothing: never a fall back to the default.
+    #[tokio::test]
+    async fn a_failed_read_keeps_the_last_value() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let path = dir.path().join("p.json");
+        let revoked = || Ok(Read::Current(record(r#"{"capabilities":[]}"#)));
+        // An empty list is atlas's default already, so only ruth changes.
+        assert_eq!(
+            sync(&path, |_| revoked()).await.expect("read"),
+            vec!["ruth".to_owned()]
+        );
+        let before = std::fs::read(&path).expect("stored");
+
+        sync(&path, |agent| {
+            if agent == "ruth" {
+                Err(anyhow::anyhow!("timed out"))
+            } else {
+                Ok(Read::Current(AgentPermissions::default()))
+            }
+        })
+        .await
+        .expect_err("a failed read fails the pass");
+        assert_eq!(std::fs::read(&path).expect("stored"), before);
+
+        let changed = sync(&path, |_| Ok(Read::NoBucket))
+            .await
+            .expect("no bucket");
+        assert!(changed.is_empty());
+        assert_eq!(std::fs::read(&path).expect("stored"), before);
+    }
+}
diff --git a/hive-host-sock/src/jobs.rs b/hive-host-sock/src/jobs.rs
index 36f7048a..189b4c69 100644
--- a/hive-host-sock/src/jobs.rs
+++ b/hive-host-sock/src/jobs.rs
@@ -25,15 +25,6 @@ pub use hive_jobq::State;
 pub enum PermPayload {
     /// Set the tool groups for one agent (`tool-groups.json`).
     ToolGroups { groups: Vec<String> },
-    /// Set the capabilities for one agent (`capabilities.json`).
-    Capabilities { caps: Vec<String> },
-    /// Set both perm-types in one entry — the batch
-    /// `POST /api/permissions` path. `None` leaves that file untouched;
-    /// present fields commit together and rebuild once.
-    Combined {
-        groups: Option<Vec<String>>,
-        caps: Option<Vec<String>>,
-    },
 }
 
 /// Node id. Carries the scheduler crate's globally-monotonic node id
diff --git a/hive-sh4re/src/permissions.rs b/hive-sh4re/src/permissions.rs
index 88247574..8713576e 100644
--- a/hive-sh4re/src/permissions.rs
+++ b/hive-sh4re/src/permissions.rs
@@ -1,8 +1,9 @@
 //! Per-agent authorization: named MCP-tool groups (`ToolGroup`) and
-//! system-level capability grants (`Capability`). Both are declared in
-//! per-agent config (`tool-groups.json` / `capabilities.json`), injected
-//! into the container as env vars, and read by the harness to decide
-//! which MCP tools claude actually sees.
+//! system-level capability grants (`Capability`). Tool groups are declared
+//! in the hive's `tool-groups.json`, injected into the container as an env
+//! var, and read by the harness to decide which MCP tools claude actually
+//! sees. Capabilities are granted in the swarm's `agent-permissions` bucket
+//! and enforced by hive-c0re.
 //!
 //! It also resolves the *built-in* claude tool list ([`builtin_tools_arg`]):
 //! that set is group-gated too (`web_tools`), so it belongs next to the
@@ -154,11 +155,11 @@ pub enum ToolGroup {
     Inbox,
     /// Gates no tool today — `list_containers`, its only member, was
     /// removed with no replacement. Kept so existing
-    /// `meta/capabilities.json` grants still parse; `tools()` returns
+    /// `meta/tool-groups.json` grants still parse; `tools()` returns
     /// `&[]`.
     Lifecycle,
     /// Gates no tool today — `create_repo`, its only member, was removed
-    /// with no replacement. Kept so existing `meta/capabilities.json`
+    /// with no replacement. Kept so existing `meta/tool-groups.json`
     /// grants still parse; `tools()` returns `&[]`.
     Forge,
     /// Gates whether the `bash` MCP server (`mcp__bash__run`/`status`/
@@ -323,16 +324,12 @@ impl ToolGroup {
     }
 }
 
-/// Per-agent capability grants. Stored in `meta/capabilities.json`
-/// (same shape as `tool-groups.json`: `{ "alice": ["manage_root_agent"] }`).
-/// Capabilities control system-level access hive-c0re enforces at
-/// dispatch time; they are orthogonal to tool groups (which control
-/// which MCP tools the harness exposes to claude).
-///
-/// Injected into containers as `HIVE_CAPABILITIES` (comma-separated
-/// `snake_case`) via `meta::render_flake`. The harness reads this to
-/// conditionally register capability-gated MCP tools so claude only
-/// sees tools it can actually invoke. See `docs/process/conventions.md::Capabilities`.
+/// Per-agent capability grants: the `capabilities` field of the swarm's
+/// `agent-permissions` record (`swarm_queue_client::agent_permissions`), set
+/// with `swarmctl agent permissions set`. Capabilities control system-level
+/// access hive-c0re enforces, through the container's bind mounts; they are
+/// orthogonal to tool groups (which control which MCP tools the harness
+/// exposes to claude). See `docs/process/conventions.md::Capabilities`.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize, strum::IntoStaticStr)]
 #[serde(rename_all = "snake_case")]
 #[strum(serialize_all = "snake_case")]
@@ -348,26 +345,14 @@ pub enum Capability {
     /// there was. With the hierarchy removed there is no privileged root
     /// to single out, so the same grant now means "manage any agent".
     /// Kept under the old spelling deliberately — renaming it would turn
-    /// every `manage_root_agent` already stored in `capabilities.json`
-    /// into an unrecognised name that `capabilities::prune_unknown`
-    /// silently drops.
+    /// every `manage_root_agent` already stored in a swarm record into a
+    /// name the hive can no longer decode.
     ManageRootAgent,
 }
 
 impl Capability {
-    /// Every known capability in a stable order. Use this to enumerate
-    /// columns in the permissions UI or validate incoming capability strings.
+    /// Every known capability in a stable order.
     pub const ALL: &'static [Self] = &[Self::ManageRootAgent];
-
-    /// Short human-readable description suitable for a tooltip or help text.
-    #[must_use]
-    pub fn description(self) -> &'static str {
-        match self {
-            Self::ManageRootAgent => {
-                "manage any agent: every agent's state (rw) and config (ro) mounted for recovery"
-            }
-        }
-    }
 }
 
 #[cfg(test)]
diff --git a/swarm-controller/Cargo.toml b/swarm-controller/Cargo.toml
index 56e7eefa..9300318e 100644
--- a/swarm-controller/Cargo.toml
+++ b/swarm-controller/Cargo.toml
@@ -66,6 +66,9 @@ sha2.workspace = true
 # `forge/pull_diff.rs` diffs a PR's changed files itself, at a context width
 # the forge's own `.diff` does not offer.
 similar.workspace = true
+# The `Capability` names `/api/agents/{name}/permissions` accepts, the same
+# enum the hive enforces.
+hive-sh4re.workspace = true
 # Validates `POST /api/agents`' `name` before it becomes `agent`/`repo`
 # everywhere downstream — see `create_agent`'s doc comment for why this is
 # defense-in-depth, not the only gate (per an argus review finding).
@@ -102,6 +105,7 @@ swarm-authelia-bridge-sock.workspace = true
 # carries.
 swarm-queue-client = { workspace = true, features = [
   "kv",
+  "agent-permissions",
   "subagent-term",
   "agent-term",
   "schema",
diff --git a/swarm-controller/src/agent_permissions.rs b/swarm-controller/src/agent_permissions.rs
new file mode 100644
index 00000000..9d8d8df1
--- /dev/null
+++ b/swarm-controller/src/agent_permissions.rs
@@ -0,0 +1,503 @@
+//! `/api/agents/{name}/permissions` — the capabilities the swarm grants one
+//! agent — and the one-shot import of the grants a hive's meta repo held.
+//!
+//! Each agent's permissions are one value under its own key of
+//! `agent-permissions` (`swarm_queue_client::agent_permissions`), which this
+//! daemon alone writes. The hive running the agent reads the key and restarts
+//! the agent when its capabilities change. A field left unset leaves the agent
+//! on the hive's compiled-in default; an empty list grants nothing.
+
+use std::collections::BTreeSet;
+use std::sync::Arc;
+
+use axum::Json;
+use axum::extract::{Path, State};
+use axum::http::StatusCode;
+use futures_util::StreamExt as _;
+use hive_sh4re::permissions::Capability;
+use problem_details::ProblemDetails;
+use serde::{Deserialize, Serialize};
+use swarm_queue_client::agent_permissions::{self, AgentPermissions, Import, ImportReply};
+use utoipa::ToSchema;
+
+use crate::schedules::{live, queue, read_or_default, refused, write};
+use crate::wanted::WantedWriter;
+use crate::{AppState, error_problem, wanted_error_status};
+
+/// How many times a losing writer re-reads and re-applies, as in
+/// `wanted::MAX_ATTEMPTS`.
+const MAX_ATTEMPTS: usize = 5;
+
+/// One agent's permissions. A field left out is not set by the swarm, and the
+/// hive's default applies; an empty list grants nothing.
+#[derive(Debug, Default, Clone, PartialEq, Eq, Serialize, Deserialize, ToSchema)]
+#[serde(deny_unknown_fields)]
+pub struct Permissions {
+    /// Capability names, `snake_case`: `manage_root_agent` mounts every agent's
+    /// state and config into this one for recovery.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[schema(value_type = Option<Vec<String>>, example = json!(["manage_root_agent"]))]
+    capabilities: Option<Vec<Capability>>,
+}
+
+impl Permissions {
+    /// The value to store, or why it is refused.
+    fn into_stored(self) -> Result<AgentPermissions, String> {
+        if let Some(caps) = &self.capabilities {
+            let mut seen = std::collections::HashSet::new();
+            for cap in caps {
+                if !seen.insert(*cap) {
+                    let name: &str = (*cap).into();
+                    return Err(format!("capability {name:?} is listed twice"));
+                }
+            }
+        }
+        Ok(AgentPermissions {
+            capabilities: self.capabilities,
+        })
+    }
+}
+
+impl From<AgentPermissions> for Permissions {
+    fn from(v: AgentPermissions) -> Self {
+        Self {
+            capabilities: v.capabilities,
+        }
+    }
+}
+
+fn agent_name(name: &str) -> Result<String, (StatusCode, String)> {
+    hive_types::Ident::parse(name)
+        .map(hive_types::Ident::into_string)
+        .map_err(|reason| (StatusCode::BAD_REQUEST, reason.to_owned()))
+}
+
+fn queue_problem(e: &anyhow::Error) -> ProblemDetails {
+    tracing::warn!(error = %format!("{e:#}"), "agent permissions: queue failure");
+    error_problem(wanted_error_status(e), &format!("{e:#}"))
+}
+
+async fn open_or_create(
+    client: &async_nats::Client,
+) -> Result<async_nats::jetstream::kv::Store, ProblemDetails> {
+    agent_permissions::open_or_create(client)
+        .await
+        .map_err(|e| queue_problem(&anyhow::Error::new(e)))
+}
+
+#[utoipa::path(
+    get,
+    path = "/api/agents/{name}/permissions",
+    params(("name" = String, Path, description = "agent name")),
+    responses(
+        (status = 200, description = "the permissions the swarm sets for the agent; `{}` \
+            when it sets none", body = Permissions),
+        (status = 400, description = "not an identifier (problem+json)", body = String),
+        (status = 503, description = "no swarm queue is configured on this host, or it is \
+            not connected (problem+json)", body = String),
+        (status = 500, description = "the permissions could not be read (problem+json)",
+            body = String),
+    ),
+    tag = "agents"
+)]
+pub(crate) async fn get_agent_permissions(
+    State(state): State<AppState>,
+    Path(name): Path<String>,
+) -> Result<Json<Permissions>, ProblemDetails> {
+    let agent = agent_name(&name).map_err(refused)?;
+    let client = queue(&state).map_err(refused)?;
+    let store = open_or_create(&client).await?;
+    let (current, _, _) = read_or_default::<AgentPermissions>(&store, &agent)
+        .await
+        .map_err(|e| queue_problem(&e))?;
+    Ok(Json(current.into()))
+}
+
+/// Set the permissions the swarm grants one agent.
+///
+/// The value replaces the stored one whole. The hive running the agent reads
+/// it within 30 seconds and, when the agent's capabilities changed, restarts
+/// it after its current turn.
+#[utoipa::path(
+    put,
+    path = "/api/agents/{name}/permissions",
+    params(("name" = String, Path, description = "agent name")),
+    request_body = Permissions,
+    responses(
+        (status = 200, description = "the permissions as stored", body = Permissions),
+        (status = 400, description = "not an identifier, not declared on any hive, or a \
+            capability listed twice (problem+json)", body = String),
+        (status = 422, description = "an unknown field or capability name"),
+        (status = 503, description = "no swarm queue is configured on this host, or it is \
+            not connected (problem+json)", body = String),
+        (status = 500, description = "the permissions could not be written (problem+json)",
+            body = String),
+    ),
+    tag = "agents"
+)]
+pub(crate) async fn put_agent_permissions(
+    State(state): State<AppState>,
+    Path(name): Path<String>,
+    Json(req): Json<Permissions>,
+) -> Result<Json<Permissions>, ProblemDetails> {
+    let agent = agent_name(&name).map_err(refused)?;
+    let value = req
+        .into_stored()
+        .map_err(|why| error_problem(StatusCode::BAD_REQUEST, &why))?;
+    let client = queue(&state).map_err(refused)?;
+    if !live(&state).await?.contains(&agent) {
+        return Err(error_problem(
+            StatusCode::BAD_REQUEST,
+            &format!("no hive's wanted state declares an agent {agent:?} that is not destroyed"),
+        ));
+    }
+    let store = open_or_create(&client).await?;
+    for _ in 0..MAX_ATTEMPTS {
+        let (current, revision, undecodable) = read_or_default::<AgentPermissions>(&store, &agent)
+            .await
+            .map_err(|e| queue_problem(&e))?;
+        if (current == value && !undecodable)
+            || write(&store, &agent, &value, revision)
+                .await
+                .map_err(|e| queue_problem(&e))?
+        {
+            tracing::info!(%agent, ?value, "set agent permissions");
+            return Ok(Json(value.into()));
+        }
+    }
+    Err(queue_problem(&anyhow::anyhow!(
+        "gave up writing the permissions of {agent} after {MAX_ATTEMPTS} conflicting writes"
+    )))
+}
+
+#[utoipa::path(
+    delete,
+    path = "/api/agents/{name}/permissions",
+    params(("name" = String, Path, description = "agent name")),
+    responses(
+        (status = 204, description = "cleared; the hive's default applies again"),
+        (status = 400, description = "not an identifier (problem+json)", body = String),
+        (status = 503, description = "no swarm queue is configured on this host, or it is \
+            not connected (problem+json)", body = String),
+        (status = 500, description = "the permissions could not be cleared (problem+json)",
+            body = String),
+    ),
+    tag = "agents"
+)]
+pub(crate) async fn delete_agent_permissions(
+    State(state): State<AppState>,
+    Path(name): Path<String>,
+) -> Result<StatusCode, ProblemDetails> {
+    let agent = agent_name(&name).map_err(refused)?;
+    let client = queue(&state).map_err(refused)?;
+    let store = open_or_create(&client).await?;
+    store
+        .delete(&agent)
+        .await
+        .map_err(|e| queue_problem(&anyhow::Error::new(e)))?;
+    tracing::info!(%agent, "cleared agent permissions");
+    Ok(StatusCode::NO_CONTENT)
+}
+
+/// Create the bucket as soon as the queue is connected, then answer every
+/// hive's import. Creating it first means a hive that finds no bucket knows
+/// the queue is not serving it, rather than that nothing is set.
+pub(crate) fn spawn(client: async_nats::Client, wanted: Arc<WantedWriter>, hives: Vec<String>) {
+    tokio::spawn(async move {
+        const RETRY: std::time::Duration = std::time::Duration::from_secs(5);
+        let store = loop {
+            if swarm_queue_client::ensure_connected(&client).is_ok() {
+                match agent_permissions::open_or_create(&client).await {
+                    Ok(store) => break store,
+                    Err(e) => tracing::warn!(
+                        error = swarm_queue_client::chain(&e),
+                        "creating the agent-permissions bucket failed; retrying"
+                    ),
+                }
+            }
+            tokio::time::sleep(RETRY).await;
+        };
+        let subject = agent_permissions::import_subject_wildcard();
+        loop {
+            match client.subscribe(subject.clone()).await {
+                Ok(mut requests) => {
+                    tracing::info!(%subject, "agent permissions: answering imports");
+                    while let Some(msg) = requests.next().await {
+                        let Some(reply) = msg.reply.clone() else {
+                            continue;
+                        };
+                        let answer = answer_import(&store, &wanted, &hives, &msg).await;
+                        let encoded = serde_json::to_vec(&answer).unwrap_or_default();
+                        if let Err(e) = client.publish(reply, encoded.into()).await {
+                            tracing::warn!(error = %e, "agent permissions: answering an import failed");
+                        }
+                    }
+                    tracing::warn!(%subject, "agent permissions: import subscription ended");
+                }
+                Err(e) => {
+                    tracing::warn!(%subject, error = %e, "agent permissions: subscribe failed");
+                }
+            }
+            tokio::time::sleep(RETRY).await;
+        }
+    });
+}
+
+async fn answer_import(
+    store: &async_nats::jetstream::kv::Store,
+    wanted: &WantedWriter,
+    hives: &[String],
+    msg: &async_nats::Message,
+) -> ImportReply {
+    let failed = |error: String| {
+        tracing::warn!(subject = %msg.subject, %error, "agent permissions: import refused");
+        ImportReply::Failed { error }
+    };
+    let Some(hive) = agent_permissions::hive_from_import_subject(&msg.subject) else {
+        return failed(format!("{} is not an import subject", msg.subject));
+    };
+    if !hives.iter().any(|h| h == hive) {
+        return failed(format!("{hive} is not a hive this controller serves"));
+    }
+    let import: Import = match serde_json::from_slice(&msg.payload) {
+        Ok(import) => import,
+        Err(e) => return failed(format!("the import does not decode: {e}")),
+    };
+    let declared = match wanted.view(hive).await {
+        Ok(d) => crate::agent_renewal::live_agents(d.as_slice()),
+        Err(e) => return failed(format!("reading {hive}'s wanted state: {e:#}")),
+    };
+    match import_into(store, &declared, import).await {
+        Ok(reply) => {
+            tracing::info!(%hive, ?reply, "agent permissions: imported a hive's grants");
+            reply
+        }
+        Err(e) => failed(format!("{e:#}")),
+    }
+}
+
+/// Write each agent's imported capabilities under its key when it has none,
+/// for the agents in `declared` only.
+async fn import_into(
+    store: &async_nats::jetstream::kv::Store,
+    declared: &BTreeSet<String>,
+    import: Import,
+) -> anyhow::Result<ImportReply> {
+    let (mut written, mut kept, mut skipped) = (Vec::new(), Vec::new(), Vec::new());
+    for (agent, caps) in import.capabilities {
+        if !declared.contains(&agent) {
+            skipped.push(agent);
+            continue;
+        }
+        let value = AgentPermissions {
+            capabilities: Some(caps),
+        };
+        if write(store, &agent, &value, None).await? {
+            written.push(agent);
+        } else {
+            kept.push(agent);
+        }
+    }
+    Ok(ImportReply::Imported {
+        written,
+        kept,
+        skipped,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{
+        Permissions, delete_agent_permissions, get_agent_permissions, import_into,
+        put_agent_permissions,
+    };
+    use crate::test_nats::TestNats;
+    use axum::extract::{Path, State};
+    use axum::http::StatusCode;
+    use hive_sh4re::permissions::Capability;
+    use swarm_queue_client::agent_permissions::{Import, ImportReply};
+
+    async fn published(client: &async_nats::Client, agent: &str) -> Option<String> {
+        let store = swarm_queue_client::agent_permissions::open_or_create(client)
+            .await
+            .expect("bucket");
+        store
+            .get(agent)
+            .await
+            .expect("read")
+            .map(|v| String::from_utf8(v.to_vec()).expect("utf-8"))
+    }
+
+    fn granted() -> Permissions {
+        Permissions {
+            capabilities: Some(vec![Capability::ManageRootAgent]),
+        }
+    }
+
+    async fn put(
+        state: &crate::AppState,
+        agent: &str,
+        value: Permissions,
+    ) -> Result<axum::Json<Permissions>, problem_details::ProblemDetails> {
+        put_agent_permissions(
+            State(state.clone()),
+            Path(agent.to_owned()),
+            axum::Json(value),
+        )
+        .await
+    }
+
+    async fn get(state: &crate::AppState, agent: &str) -> Permissions {
+        get_agent_permissions(State(state.clone()), Path(agent.to_owned()))
+            .await
+            .expect("read")
+            .0
+    }
+
+    /// Through the real handlers: a grant lands as the hive reads it, an empty
+    /// list is stored as an empty list, and `DELETE` takes the key out.
+    #[tokio::test]
+    async fn the_route_writes_what_a_hive_reads_and_delete_clears_it() {
+        let nats = TestNats::start();
+        let client = nats.client().await;
+        let state = crate::agent_limits::tests::declared(&client).await;
+
+        assert_eq!(
+            put(&state, "atlas", granted()).await.expect("set").0,
+            granted()
+        );
+        assert_eq!(
+            published(&client, "atlas").await.as_deref(),
+            Some(r#"{"capabilities":["manage_root_agent"]}"#)
+        );
+        assert_eq!(get(&state, "atlas").await, granted());
+
+        let none = Permissions {
+            capabilities: Some(Vec::new()),
+        };
+        let _ = put(&state, "atlas", none.clone()).await.expect("set");
+        assert_eq!(
+            published(&client, "atlas").await.as_deref(),
+            Some(r#"{"capabilities":[]}"#)
+        );
+
+        assert_eq!(
+            delete_agent_permissions(State(state.clone()), Path("atlas".to_owned())).await,
+            Ok(StatusCode::NO_CONTENT)
+        );
+        assert_eq!(published(&client, "atlas").await, None);
+        assert_eq!(get(&state, "atlas").await, Permissions::default());
+    }
+
+    /// A name no hive declares, or a name that is not an identifier, is a 400
+    /// and writes nothing.
+    #[tokio::test]
+    async fn an_undeclared_or_malformed_agent_is_a_400() {
+        let nats = TestNats::start();
+        let client = nats.client().await;
+        let state = crate::agent_limits::tests::declared(&client).await;
+
+        for agent in ["iris", "Not_An_Ident"] {
+            let problem = put(&state, agent, granted()).await.expect_err("refused");
+            assert_eq!(problem.status, Some(StatusCode::BAD_REQUEST), "{agent}");
+        }
+        assert_eq!(published(&client, "iris").await, None);
+    }
+
+    async fn put_raw(app: &axum::Router, body: &'static str) -> StatusCode {
+        use tower::ServiceExt as _;
+        app.clone()
+            .oneshot(
+                axum::http::Request::put("/api/agents/atlas/permissions")
+                    .header(axum::http::header::CONTENT_TYPE, "application/json")
+                    .body(axum::body::Body::from(body))
+                    .expect("request"),
+            )
+            .await
+            .expect("router answers")
+            .status()
+    }
+
+    /// Behind the real router: an unknown field or capability name is a 422,
+    /// a duplicate is a 400, and none of them changes what is stored.
+    #[tokio::test]
+    async fn a_bad_body_is_refused_and_the_grant_stays() {
+        const GRANTED: &str = r#"{"capabilities":["manage_root_agent"]}"#;
+        let nats = TestNats::start();
+        let client = nats.client().await;
+        let state = crate::agent_limits::tests::declared(&client).await;
+        let _ = put(&state, "atlas", granted()).await.expect("set");
+        let app = crate::build_app(state);
+
+        for (body, status) in [
+            (
+                r#"{"capabilities":["manage_root_agent"],"tool_groups":["meta"]}"#,
+                StatusCode::UNPROCESSABLE_ENTITY,
+            ),
+            (r#"{"capabilites":[]}"#, StatusCode::UNPROCESSABLE_ENTITY),
+            (
+                r#"{"capabilities":["fly_to_the_moon"]}"#,
+                StatusCode::UNPROCESSABLE_ENTITY,
+            ),
+            (
+                r#"{"capabilities":["manage_root_agent","manage_root_agent"]}"#,
+                StatusCode::BAD_REQUEST,
+            ),
+        ] {
+            assert_eq!(put_raw(&app, body).await, status, "{body}");
+            assert_eq!(
+                published(&client, "atlas").await.as_deref(),
+                Some(GRANTED),
+                "{body}"
+            );
+        }
+        // Control: the same router and request do reach the bucket.
+        assert_eq!(put_raw(&app, "{}").await, StatusCode::OK);
+        assert_eq!(published(&client, "atlas").await.as_deref(), Some("{}"));
+    }
+
+    /// The import writes a key only where there is none, and only for the
+    /// agents the sending hive declares.
+    #[tokio::test]
+    async fn an_import_writes_only_missing_keys_of_the_hives_own_agents() {
+        let nats = TestNats::start();
+        let client = nats.client().await;
+        let store = swarm_queue_client::agent_permissions::open_or_create(&client)
+            .await
+            .expect("bucket");
+        store
+            .put("atlas", br#"{"capabilities":[]}"#.to_vec().into())
+            .await
+            .expect("seed");
+        let declared = ["atlas".to_owned(), "ruth".to_owned()].into();
+        let import = Import {
+            capabilities: [
+                ("atlas".to_owned(), vec![Capability::ManageRootAgent]),
+                ("ruth".to_owned(), vec![Capability::ManageRootAgent]),
+                ("iris".to_owned(), vec![Capability::ManageRootAgent]),
+            ]
+            .into(),
+        };
+
+        let reply = import_into(&store, &declared, import)
+            .await
+            .expect("imported");
+        assert_eq!(
+            reply,
+            ImportReply::Imported {
+                written: vec!["ruth".to_owned()],
+                kept: vec!["atlas".to_owned()],
+                skipped: vec!["iris".to_owned()],
+            }
+        );
+        assert_eq!(
+            published(&client, "atlas").await.as_deref(),
+            Some(r#"{"capabilities":[]}"#)
+        );
+        assert_eq!(
+            published(&client, "ruth").await.as_deref(),
+            Some(r#"{"capabilities":["manage_root_agent"]}"#)
+        );
+        assert_eq!(published(&client, "iris").await, None);
+    }
+}
diff --git a/swarm-controller/src/main.rs b/swarm-controller/src/main.rs
index dd30e4db..7d77cf00 100644
--- a/swarm-controller/src/main.rs
+++ b/swarm-controller/src/main.rs
@@ -47,6 +47,7 @@ mod agent_identity;
 mod agent_input;
 mod agent_limits;
 mod agent_login;
+mod agent_permissions;
 mod agent_renewal;
 mod agent_state_stream;
 mod agent_status;
@@ -2877,11 +2878,22 @@ fn spawn_subagent_streams(status: Option<&Arc<status::StatusReader>>, hives: &[H
 }
 
 /// The per-agent background passes: subagent and terminal streams, and
-/// schedule keys.
+/// schedule keys, and the permissions bucket and its import.
 /// Lifted out of `main` for `clippy::too_many_lines`.
 fn spawn_per_agent_upkeep(status: Option<&Arc<status::StatusReader>>, hives: &[HiveEntry]) {
     spawn_subagent_streams(status, hives);
     spawn_schedule_reconcile(status, hives);
+    spawn_agent_permissions(status, hives);
+}
+
+/// Create the agent-permissions bucket and answer hives' imports when a swarm
+/// queue is wired up. Same shape as [`spawn_subagent_streams`].
+fn spawn_agent_permissions(status: Option<&Arc<status::StatusReader>>, hives: &[HiveEntry]) {
+    let (Some(status), Some(wanted)) = (status, wanted_writer(status)) else {
+        return;
+    };
+    let hives = hives.iter().map(|h| h.name.clone()).collect();
+    agent_permissions::spawn(status.queue_client(), wanted, hives);
 }
 
 /// Start rewriting every agent's schedule key from the definitions when a
@@ -3493,6 +3505,11 @@ fn build_app(state: AppState) -> axum::Router {
             agent_limits::put_agent_limits,
             agent_limits::delete_agent_limits
         ))
+        .routes(routes!(
+            agent_permissions::get_agent_permissions,
+            agent_permissions::put_agent_permissions,
+            agent_permissions::delete_agent_permissions
+        ))
         .routes(routes!(get_config_prs))
         .routes(routes!(create_agent))
         .routes(routes!(mint_agent_forge_token))
diff --git a/swarm-nats-auth/src/policy.rs b/swarm-nats-auth/src/policy.rs
index 94463a40..cf3ab2d7 100644
--- a/swarm-nats-auth/src/policy.rs
+++ b/swarm-nats-auth/src/policy.rs
@@ -229,7 +229,9 @@ impl Policy {
             return Some(Permissions {
                 publish: self.reader_subjects(),
                 subscribe: Some(Self::reader_subscribe(client_id)),
-                responses: None,
+                // The answer to a hive's permissions import. NATS lets it
+                // answer only a message it received, once.
+                responses: Some(swarm_queue_client::agent_permissions::IMPORT_REPLY_DEADLINE),
             });
         }
         if self.review_classifiers.iter().any(|c| c == client_id) {
@@ -307,7 +309,7 @@ impl Policy {
 
     /// What a reader may subscribe to: every agent's terminal, its
     /// subagents' terminals and its turn state, which the controller relays
-    /// to swarm-ui, and its own reply inbox.
+    /// to swarm-ui, every hive's permissions import, and its own reply inbox.
     ///
     /// The three families are the publishers' (`hive-agent`'s `swarm_term`
     /// and `swarm_agent_state`, `swarm_queue_client::subagent_term`), spelled
@@ -318,6 +320,7 @@ impl Policy {
             swarm_queue_client::term_subject_wildcard(),
             swarm_queue_client::subagent_term::subject_wildcard(),
             swarm_queue_client::agent_state_subject_wildcard(),
+            swarm_queue_client::agent_permissions::import_subject_wildcard(),
             Self::inbox(client_id),
         ]
     }
@@ -561,6 +564,17 @@ impl Policy {
             "$JS.API.DIRECT.GET.{}.>",
             swarm_queue_client::agent_limits::STREAM
         ));
+        // The capabilities the swarm grants those agents
+        // (`swarm_queue_client::agent_permissions`), read the same way and
+        // every key for the same reason.
+        subjects.push(format!(
+            "$JS.API.DIRECT.GET.{}.>",
+            swarm_queue_client::agent_permissions::STREAM
+        ));
+        // The one-shot import of the grants this hive's meta repo held, sent
+        // as a request on this hive's own subject: the controller takes the
+        // hive from the subject, so one hive cannot import for another.
+        subjects.push(swarm_queue_client::agent_permissions::import_subject(hive));
         subjects.extend(
             self.extra_hive_subjects
                 .iter()
@@ -837,6 +851,78 @@ mod tests {
         }
     }
 
+    /// A hive reads any agent's permissions and can change none of them; an
+    /// agent, even one granted direct gets on its own keys elsewhere, reads
+    /// none of them.
+    #[test]
+    fn a_hive_may_read_every_agents_permissions_and_an_agent_none() {
+        use swarm_queue_client::agent_permissions::{STREAM, direct_get_subject, key_subject};
+
+        let p = policy()
+            .with_agent_token_subjects(vec![
+                "$JS.API.DIRECT.GET.KV_agent-inference.$KV.agent-inference.{agent}".to_owned(),
+            ])
+            .expect("a per-agent template is valid");
+        let hive = p.permissions("hive-alpha").expect("a hive is admitted");
+        let hive_may = |subject: &str| hive.publish.iter().any(|g| covers(g, subject));
+        for agent in ["atlas", "ruth"] {
+            assert!(hive_may(&direct_get_subject(agent)), "cannot read {agent}");
+        }
+        for forbidden in [
+            key_subject("atlas"),
+            format!("$JS.API.STREAM.CREATE.{STREAM}"),
+            format!("$JS.API.STREAM.DELETE.{STREAM}"),
+            format!("$JS.API.STREAM.PURGE.{STREAM}"),
+            format!("$JS.API.CONSUMER.CREATE.{STREAM}"),
+        ] {
+            assert!(!hive_may(&forbidden), "hive alpha may publish {forbidden}");
+        }
+
+        let agent = p.agent_token_permissions("atlas").expect("configured");
+        let agent_may = |subject: &str| agent.publish.iter().any(|g| covers(g, subject));
+        assert!(
+            !agent_may(&direct_get_subject("atlas")),
+            "an agent may read its own permissions: {:?}",
+            agent.publish
+        );
+        // Control: the agent's direct-get grant is real, just not this bucket.
+        assert!(agent_may(
+            &swarm_queue_client::agent_inference::direct_get_subject("atlas")
+        ));
+    }
+
+    /// A hive sends its import on its own subject only; the reader hears every
+    /// hive's and may answer, and no other role may send one.
+    #[test]
+    fn a_hive_imports_only_under_its_own_name_and_the_reader_answers() {
+        use swarm_queue_client::agent_permissions::{
+            IMPORT_REPLY_DEADLINE, import_subject, import_subject_wildcard,
+        };
+
+        let p = policy();
+        let hive = p.permissions("hive-alpha").expect("a hive is admitted");
+        let hive_may = |subject: &str| hive.publish.iter().any(|g| covers(g, subject));
+        assert!(hive_may(&import_subject("alpha")));
+        assert!(!hive_may(&import_subject("beta")));
+
+        let reader = p.permissions("swarm-controller").expect("the reader");
+        assert!(
+            reader
+                .subscribe
+                .as_ref()
+                .expect("scoped")
+                .contains(&import_subject_wildcard())
+        );
+        assert_eq!(reader.responses, Some(IMPORT_REPLY_DEADLINE));
+        assert!(
+            !reader
+                .publish
+                .iter()
+                .any(|g| covers(g, &import_subject("alpha"))),
+            "the reader may send an import as a hive"
+        );
+    }
+
     #[test]
     fn an_extra_subject_without_the_placeholder_is_refused() {
         // 🩸 The option exists to put a second stream inside ONE hive's
@@ -1356,7 +1442,6 @@ mod tests {
             .expect("a per-agent template is valid");
         let reader = p.permissions("swarm-controller").expect("the reader");
         assert!(reader.publish.iter().any(|g| covers(g, SUBJECT)));
-        assert_eq!(reader.responses, None);
         for (who, grant) in [
             ("a hive", p.permissions("hive-alpha").expect("a hive")),
             (
diff --git a/swarm-queue-client/Cargo.toml b/swarm-queue-client/Cargo.toml
index 0fb0389f..4ce7b27f 100644
--- a/swarm-queue-client/Cargo.toml
+++ b/swarm-queue-client/Cargo.toml
@@ -50,6 +50,10 @@ schema = ["dep:utoipa"]
 # `agent_limits::read`, for hive-c0re, which reads the key of each agent it
 # runs the same way. Explicit for the same reason as `notices`.
 agent-limits = ["async-nats/jetstream"]
+# `agent_permissions::AgentPermissions` and `read`, for hive-c0re and
+# swarm-controller. Pulls `hive-sh4re` for the `Capability` names, so the
+# crates that only need the bucket name do not link it.
+agent-permissions = ["dep:hive-sh4re", "async-nats/jetstream"]
 
 [dependencies]
 # Bare (no `kv`/`jetstream`) unless a consumer opts into the `kv` feature
@@ -58,6 +62,7 @@ async-nats.workspace = true
 # `Shared`, so every caller refused while a secret re-read runs waits on that
 # one read (`stored_secret`).
 futures-util.workspace = true
+hive-sh4re = { workspace = true, optional = true }
 # `blocking` on top of the workspace default (`form`/`json`/`rustls`) —
 # `mint_token_for_blocking` needs `reqwest::blocking::Client` for a caller
 # with no tokio reactor to `.await` an async request on (an OTLP exporter's
diff --git a/swarm-queue-client/src/agent_permissions.rs b/swarm-queue-client/src/agent_permissions.rs
new file mode 100644
index 00000000..e712eaac
--- /dev/null
+++ b/swarm-queue-client/src/agent_permissions.rs
@@ -0,0 +1,255 @@
+//! The per-agent permissions KV bucket: one key per agent, holding the
+//! capabilities the swarm grants it.
+//!
+//! **swarm-controller is the only writer**, and creates the bucket when it
+//! starts, so a bucket that does not answer means the queue is not serving it,
+//! never "nothing set". A field left unset leaves the agent on its hive's
+//! compiled-in default; an empty list grants nothing.
+//!
+//! **The hive running the agent reads the key**, not the agent: capabilities
+//! are bind mounts the host sets up at container start. A hive reads with one
+//! `DIRECT.GET` per agent
+//! ([`crate::agent_permissions::direct_get_subject`]), and its grant covers
+//! every key of this bucket, read-only, for the reason [`crate::agent_limits`]
+//! gives.
+//!
+//! Keyed by the agent alone, like [`crate::agent_limits`]: an agent can move
+//! between hives, and its permissions go with it.
+//!
+//! Also carries the one-shot import a hive sends of the grants its meta repo
+//! held ([`crate::agent_permissions::import_subject`]), answered by the
+//! controller.
+
+/// The KV bucket permissions are published into, keyed by agent name.
+pub const BUCKET: &str = "agent-permissions";
+
+/// The stream backing [`BUCKET`]. `create_key_value` prefixes `KV_`.
+pub const STREAM: &str = "KV_agent-permissions";
+
+/// The subject `agent`'s entry is stored under.
+#[must_use]
+pub fn key_subject(agent: &str) -> String {
+    format!("$KV.{BUCKET}.{agent}")
+}
+
+/// The request subject that reads `agent`'s current entry.
+#[must_use]
+pub fn direct_get_subject(agent: &str) -> String {
+    format!("$JS.API.DIRECT.GET.{STREAM}.{}", key_subject(agent))
+}
+
+const IMPORT_SUBJECT_PREFIX: &str = "$SWARM.permissions-import";
+
+/// The subject `hive` sends its import on. One per hive, so the hive's grant
+/// names only its own, and the controller takes the hive from the subject
+/// rather than from the payload.
+#[must_use]
+pub fn import_subject(hive: &str) -> String {
+    format!("{IMPORT_SUBJECT_PREFIX}.{hive}")
+}
+
+/// Every [`import_subject`] at once, for the controller's subscription.
+#[must_use]
+pub fn import_subject_wildcard() -> String {
+    import_subject("*")
+}
+
+/// The hive an [`import_subject`] names, or `None` for any other subject.
+#[must_use]
+pub fn hive_from_import_subject(subject: &str) -> Option<&str> {
+    let hive = subject
+        .strip_prefix(IMPORT_SUBJECT_PREFIX)?
+        .strip_prefix('.')?;
+    (!hive.is_empty() && !hive.contains('.')).then_some(hive)
+}
+
+/// How long a hive waits for the controller's answer to an import, and the
+/// queue lets the controller answer.
+pub const IMPORT_REPLY_DEADLINE: std::time::Duration = std::time::Duration::from_secs(30);
+
+/// One agent's permissions: the value under `BUCKET`/`<agent>`. `None` is
+/// "not set by the swarm", which a hive reads as its default.
+///
+/// No `deny_unknown_fields`: a field a newer controller adds must not cost an
+/// older hive the fields it knows.
+#[cfg(feature = "agent-permissions")]
+#[derive(Debug, Clone, Default, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+pub struct AgentPermissions {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub capabilities: Option<Vec<hive_sh4re::permissions::Capability>>,
+}
+
+/// What a hive sends once: the capabilities its meta repo held, by agent.
+#[cfg(feature = "agent-permissions")]
+#[derive(Debug, Clone, Default, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+pub struct Import {
+    pub capabilities: std::collections::BTreeMap<String, Vec<hive_sh4re::permissions::Capability>>,
+}
+
+/// The controller's answer to an [`Import`].
+#[cfg(feature = "agent-permissions")]
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(tag = "state", rename_all = "kebab-case")]
+pub enum ImportReply {
+    /// Every agent was handled. `written` had no key and got the hive's value;
+    /// `kept` already had a key, which was left alone; `skipped` is not an
+    /// agent the sending hive's wanted state declares.
+    Imported {
+        written: Vec<String>,
+        kept: Vec<String>,
+        skipped: Vec<String>,
+    },
+    /// Nothing is known to be written; the hive sends again later.
+    Failed { error: String },
+}
+
+/// Open the permissions bucket, creating it if nothing has yet.
+///
+/// **Controller-side only.** `history: 1`: a hive wants the current value.
+#[cfg(feature = "kv")]
+pub async fn open_or_create(
+    client: &async_nats::Client,
+) -> Result<async_nats::jetstream::kv::Store, crate::Error> {
+    let js = async_nats::jetstream::new(client.clone());
+    crate::kv::open_or_create(
+        &js,
+        BUCKET,
+        "agent-permissions",
+        "Permissions the swarm controller sets for each agent".to_owned(),
+    )
+    .await
+}
+
+/// What reading one agent's key found.
+#[cfg(feature = "agent-permissions")]
+#[derive(Debug)]
+pub enum Read {
+    /// The bucket answered; an absent or deleted key has every field unset.
+    Current(AgentPermissions),
+    /// Nothing answers for the bucket's stream. Not an unset value: it says
+    /// nothing about the agent's permissions.
+    NoBucket,
+}
+
+/// Read `agent`'s current permissions with one direct get.
+///
+/// # Errors
+///
+/// A failed request or an undecodable value. A request the queue refused for
+/// permissions reaches the client as a timeout, not as a refusal.
+#[cfg(feature = "agent-permissions")]
+pub async fn read(
+    client: &async_nats::Client,
+    agent: &str,
+) -> Result<Read, Box<dyn std::error::Error + Send + Sync>> {
+    use async_nats::jetstream::stream::DirectGetErrorKind;
+
+    let js = async_nats::jetstream::new(client.clone());
+    let stream = js.get_stream_no_info(STREAM).await?;
+    let message = match stream.direct_get_last_for_subject(key_subject(agent)).await {
+        Ok(message) => message,
+        Err(e) if matches!(e.kind(), DirectGetErrorKind::NotFound) => {
+            return Ok(Read::Current(AgentPermissions::default()));
+        }
+        Err(e)
+            if matches!(
+                e.kind(),
+                DirectGetErrorKind::ErrorResponse(status, _)
+                    if status == async_nats::StatusCode::NO_RESPONDERS
+            ) =>
+        {
+            return Ok(Read::NoBucket);
+        }
+        Err(e) => return Err(e.into()),
+    };
+    Ok(Read::Current(decode(
+        message
+            .headers
+            .get("KV-Operation")
+            .map(async_nats::HeaderValue::as_str),
+        &message.payload,
+    )?))
+}
+
+/// One stored message as a value: a KV delete or purge is stored as an empty
+/// marker carrying the `KV-Operation` header, and reads as every field unset.
+#[cfg(feature = "agent-permissions")]
+fn decode(operation: Option<&str>, payload: &[u8]) -> serde_json::Result<AgentPermissions> {
+    if matches!(operation, Some("DEL" | "PURGE")) {
+        return Ok(AgentPermissions::default());
+    }
+    serde_json::from_slice(payload)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The names the grants in `swarm-nats-auth` are built from.
+    #[test]
+    fn names_match_the_queue_grant() {
+        assert_eq!(STREAM, format!("KV_{BUCKET}"));
+        assert_eq!(key_subject("iris"), "$KV.agent-permissions.iris");
+        assert_eq!(
+            direct_get_subject("iris"),
+            "$JS.API.DIRECT.GET.KV_agent-permissions.$KV.agent-permissions.iris"
+        );
+        assert_eq!(import_subject("pr1ma"), "$SWARM.permissions-import.pr1ma");
+        assert_eq!(import_subject_wildcard(), "$SWARM.permissions-import.*");
+    }
+
+    #[test]
+    fn the_hive_is_read_back_from_its_import_subject_only() {
+        assert_eq!(
+            hive_from_import_subject(&import_subject("pr1ma")),
+            Some("pr1ma")
+        );
+        for other in [
+            "$SWARM.permissions-import",
+            "$SWARM.permissions-import.",
+            "$SWARM.permissions-import.a.b",
+            "$SWARM.permissions-importx.pr1ma",
+            "$SWARM.deploy.pr1ma",
+        ] {
+            assert_eq!(hive_from_import_subject(other), None, "{other}");
+        }
+    }
+
+    /// Unset is left out on the wire, `{}` is unset, and an empty list stays
+    /// an empty list: "grants nothing" is not "not set".
+    #[cfg(feature = "agent-permissions")]
+    #[test]
+    fn unset_and_empty_are_different_values() {
+        use hive_sh4re::permissions::Capability;
+        assert_eq!(
+            serde_json::to_string(&AgentPermissions::default()).unwrap(),
+            "{}"
+        );
+        let none: AgentPermissions = serde_json::from_str(r#"{"capabilities":[]}"#).unwrap();
+        assert_eq!(none.capabilities, Some(Vec::new()));
+        let ruth: AgentPermissions =
+            serde_json::from_str(r#"{"capabilities":["manage_root_agent"]}"#).unwrap();
+        assert_eq!(ruth.capabilities, Some(vec![Capability::ManageRootAgent]));
+    }
+
+    /// Fields a newer controller adds must not cost an older hive the ones it
+    /// knows; a deleted key is unset.
+    #[cfg(feature = "agent-permissions")]
+    #[test]
+    fn decoding_tolerates_unknown_fields_and_reads_a_delete_as_unset() {
+        let decoded = decode(
+            None,
+            br#"{"capabilities":["manage_root_agent"],"tool_groups":["meta"]}"#,
+        )
+        .unwrap();
+        assert_eq!(decoded.capabilities.map(|c| c.len()), Some(1));
+        assert_eq!(
+            decode(Some("DEL"), b"").unwrap(),
+            AgentPermissions::default()
+        );
+        assert_eq!(
+            decode(Some("PURGE"), b"").unwrap(),
+            AgentPermissions::default()
+        );
+    }
+}
diff --git a/swarm-queue-client/src/lib.rs b/swarm-queue-client/src/lib.rs
index b2b31dc2..a064af7e 100644
--- a/swarm-queue-client/src/lib.rs
+++ b/swarm-queue-client/src/lib.rs
@@ -221,6 +221,12 @@ pub mod agent_inference;
 /// writing needs `kv` and reading needs `agent-limits`.
 pub mod agent_limits;
 
+/// The bucket swarm-controller sets each agent's permissions in — one key
+/// per agent, read by the hive that runs it, plus the one-shot import a hive
+/// sends. The names are unconditional; writing needs `kv` and the stored type
+/// and reading need `agent-permissions`.
+pub mod agent_permissions;
+
 /// The login flag in an agent's turn-state header. No bucket: it travels
 /// on [`agent_state_subject`] only.
 pub mod agent_login;
diff --git a/swarmctl/src/main.rs b/swarmctl/src/main.rs
index 92f802fe..10346ceb 100644
--- a/swarmctl/src/main.rs
+++ b/swarmctl/src/main.rs
@@ -29,6 +29,7 @@ mod agent;
 mod forge;
 mod inference;
 mod login;
+mod permissions;
 mod remote;
 mod schedule;
 mod users;
@@ -290,6 +291,43 @@ enum AgentVerb {
         #[command(subcommand)]
         command: InferenceVerb,
     },
+    /// Show or change the capabilities the swarm grants an agent.
+    ///
+    /// The hive running the agent reads them every 30 seconds and, when they
+    /// change, restarts the agent after its current turn. Unset leaves the
+    /// agent on its hive's default: none, except `ruth`, who gets
+    /// `manage_root_agent`.
+    Permissions {
+        #[command(subcommand)]
+        command: PermissionsVerb,
+    },
+}
+
+#[derive(Subcommand)]
+enum PermissionsVerb {
+    /// Show the capabilities the swarm grants an agent, or that it sets none.
+    Get(InferenceAgentArgs),
+    /// Replace the capabilities the swarm grants an agent.
+    Set(PermissionsSetArgs),
+    /// Clear the agent's permissions, so its hive's default applies again.
+    Clear(InferenceAgentArgs),
+}
+
+#[derive(Args)]
+struct PermissionsSetArgs {
+    /// Agent name. A hive must declare it.
+    agent: String,
+    /// The capabilities to grant, comma-separated (`manage_root_agent`).
+    /// `--capabilities ''` grants none.
+    #[arg(
+        long,
+        value_name = "CAPABILITIES",
+        value_delimiter = ',',
+        required = true
+    )]
+    capabilities: Vec<String>,
+    #[command(flatten)]
+    socket: ControllerSocketArg,
 }
 
 #[derive(Subcommand)]
@@ -602,6 +640,9 @@ fn main() -> Result<()> {
         Verb::Agent {
             command: AgentVerb::Inference { command },
         } => inference(&remote, command),
+        Verb::Agent {
+            command: AgentVerb::Permissions { command },
+        } => permissions(&remote, command),
         // Same socket-resolution reasoning as `Create` above.
         Verb::Forge {
             command: ForgeVerb::MakeAdmin(args),
@@ -671,6 +712,16 @@ fn inference(remote: &RemoteArgs, command: InferenceVerb) -> Result<()> {
     }
 }
 
+fn permissions(remote: &RemoteArgs, command: PermissionsVerb) -> Result<()> {
+    match command {
+        PermissionsVerb::Get(a) => permissions::get(&a.socket.resolve(remote)?, &a.agent),
+        PermissionsVerb::Set(a) => {
+            permissions::set(&a.socket.resolve(remote)?, &a.agent, a.capabilities)
+        }
+        PermissionsVerb::Clear(a) => permissions::clear(&a.socket.resolve(remote)?, &a.agent),
+    }
+}
+
 fn schedule(remote: &RemoteArgs, command: ScheduleVerb) -> Result<()> {
     match command {
         ScheduleVerb::Add(a) => schedule::add(
diff --git a/swarmctl/src/permissions.rs b/swarmctl/src/permissions.rs
new file mode 100644
index 00000000..67e2ef2b
--- /dev/null
+++ b/swarmctl/src/permissions.rs
@@ -0,0 +1,126 @@
+//! `swarmctl agent permissions` — the capabilities the swarm grants one
+//! agent, through swarm-controller's `/api/agents/{name}/permissions` routes.
+//!
+//! The hive running the agent reads its key every 30 seconds and restarts the
+//! agent, after its current turn, when its capabilities change. Unset leaves
+//! the agent on its hive's default; an empty list grants nothing.
+//!
+//! The shape below mirrors the controller's `Permissions`, for the reason
+//! [`crate::agent`]'s module doc gives. Capability names are checked by the
+//! controller, which answers 422 for one it does not know.
+
+use anyhow::{Context as _, Result};
+use serde::{Deserialize, Serialize};
+
+use crate::agent::{parse_ident, send};
+use crate::remote::Target;
+
+/// One agent's permissions; an absent field is not set at swarm level.
+#[derive(Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
+struct Permissions {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    capabilities: Option<Vec<String>>,
+}
+
+fn runtime() -> Result<tokio::runtime::Runtime> {
+    tokio::runtime::Builder::new_current_thread()
+        .enable_all()
+        .build()
+        .context("starting a tokio runtime for the controller request")
+}
+
+/// The agent's permissions path. The name is checked first, so it cannot
+/// reshape the URI.
+fn member(agent: &str) -> Result<String> {
+    let agent = parse_ident(agent, "agent name")?;
+    Ok(format!("/api/agents/{agent}/permissions"))
+}
+
+/// Run `swarmctl agent permissions get`.
+pub(crate) fn get(target: &Target, agent: &str) -> Result<()> {
+    let uri = member(agent)?;
+    let stored: Permissions = runtime()?.block_on(send(
+        target,
+        hyper::Method::GET,
+        &uri,
+        None::<&()>,
+        "permissions-get",
+    ))?;
+    print!("{}", render(&stored));
+    Ok(())
+}
+
+/// Run `swarmctl agent permissions set`: store the given capabilities, which
+/// replace the agent's record whole.
+pub(crate) fn set(target: &Target, agent: &str, capabilities: Vec<String>) -> Result<()> {
+    let uri = member(agent)?;
+    let stored: Permissions = runtime()?.block_on(send(
+        target,
+        hyper::Method::PUT,
+        &uri,
+        Some(&with_capabilities(capabilities)),
+        "permissions-set",
+    ))?;
+    print!("{}", render(&stored));
+    Ok(())
+}
+
+/// The record granting `capabilities`. An empty name, as `--capabilities ''`
+/// gives, is dropped, so that spelling grants nothing.
+fn with_capabilities(capabilities: Vec<String>) -> Permissions {
+    Permissions {
+        capabilities: Some(capabilities.into_iter().filter(|c| !c.is_empty()).collect()),
+    }
+}
+
+/// Run `swarmctl agent permissions clear`.
+pub(crate) fn clear(target: &Target, agent: &str) -> Result<()> {
+    let uri = member(agent)?;
+    runtime()?.block_on(send::<(), ()>(
+        target,
+        hyper::Method::DELETE,
+        &uri,
+        None,
+        "permissions-clear",
+    ))?;
+    println!("cleared; {agent} goes back to its hive's default");
+    Ok(())
+}
+
+fn render(p: &Permissions) -> String {
+    let caps = match p.capabilities.as_deref() {
+        None => "(not set; the hive's default)".to_owned(),
+        Some([]) => "(none)".to_owned(),
+        Some(caps) => caps.join(", "),
+    };
+    format!("capabilities: {caps}\n")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn get_tells_unset_from_none() {
+        let unset: Permissions = serde_json::from_str("{}").unwrap();
+        assert_eq!(
+            render(&unset),
+            "capabilities: (not set; the hive's default)\n"
+        );
+        let none: Permissions = serde_json::from_str(r#"{"capabilities":[]}"#).unwrap();
+        assert_eq!(render(&none), "capabilities: (none)\n");
+        let ruth: Permissions =
+            serde_json::from_str(r#"{"capabilities":["manage_root_agent"]}"#).unwrap();
+        assert_eq!(render(&ruth), "capabilities: manage_root_agent\n");
+    }
+
+    /// `--capabilities ''` stores an empty list, not one empty name.
+    #[test]
+    fn an_empty_list_flag_grants_nothing() {
+        let set = with_capabilities(vec![String::new()]);
+        assert_eq!(
+            serde_json::to_string(&set).unwrap(),
+            r#"{"capabilities":[]}"#
+        );
+    }
+}