diff --git a/TODO.md b/TODO.md
index 3cb7332..fc1d8c4 100644
--- a/TODO.md
+++ b/TODO.md
@@ -3,17 +3,6 @@
 Pick anything from here when relevant. Cross-cutting design notes live in
 [CLAUDE.md](CLAUDE.md); high-level project intro in [README.md](README.md).
 
-## Permissions / policy
-
-- **Per-agent send allow-list.** Today any agent can `send` to any
-  other recipient (peer, manager, operator). Add a per-agent
-  policy that constrains the `to` field — declared in `agent.nix`,
-  e.g. `hyperhive.allowedRecipients = [ "manager" "alice" ]`.
-  Broker rejects with an `Err { message }` when the policy denies.
-  Default: unrestricted (back-compat). The manager can still
-  always send anywhere. Useful for sandboxing untrusted sub-agents
-  so they can only talk to the manager, not other sub-agents.
-
 ## Security
 
 - **Unprivileged containers (userns mapping).** Today the nspawn container
diff --git a/hive-ag3nt/prompts/agent.md b/hive-ag3nt/prompts/agent.md
index 88a2d69..094b30d 100644
--- a/hive-ag3nt/prompts/agent.md
+++ b/hive-ag3nt/prompts/agent.md
@@ -3,7 +3,7 @@ You are hyperhive agent `{label}` in a multi-agent system. The operator (recipie
 Tools (hyperhive surface):
 
 - `mcp__hyperhive__recv(wait_seconds?)` — drain one more message from your inbox (returns `(empty)` if nothing pending). Without `wait_seconds` (or with `0`) it returns immediately — a cheap "anything pending?" peek you can sprinkle between tool calls. To **wait** for work when you have nothing else useful to do this turn, call with a long wait (e.g. `wait_seconds: 180`, the max) — incoming messages wake you instantly, otherwise the call returns empty at the timeout. That's strictly better than a fixed `sleep` shell command: lower latency on new work, no busy-loop.
-- `mcp__hyperhive__send(to, body)` — message a peer (by their name) or the operator (recipient `operator`, surfaces in the dashboard).
+- `mcp__hyperhive__send(to, body)` — message a peer (by their name) or the operator (recipient `operator`, surfaces in the dashboard). Some agents have a per-agent allow-list (`hyperhive.allowedRecipients` in their `agent.nix`) — if so the tool refuses recipients outside the list with a clear error; route through the manager (`send(to: "manager", …)`) which is always reachable.
 - (some agents only) **extra MCP tools** surfaced as `mcp__<server>__<tool>` — these are agent-specific (matrix client, scraper, db connector, etc.) declared in your `agent.nix` under `hyperhive.extraMcpServers`. Treat them as first-class tools alongside the hyperhive surface; the operator already auto-approved them at deploy time.
 - `mcp__hyperhive__ask_operator(question, options?, multi?, ttl_seconds?)` — surface a question to the human operator on the dashboard. Returns immediately with a question id — do NOT wait inline. When the operator answers, a system message with event `operator_answered { id, question, answer }` lands in your inbox; handle it on a future turn. Use this for clarifications, permission for risky actions, or choice between options. `options` is advisory: a short fixed-choice list when applicable, otherwise leave empty for free text. `multi: true` lets the operator pick multiple (checkboxes), answer comes back comma-joined. `ttl_seconds` auto-cancels with answer `[expired]` when the decision becomes moot.
 
diff --git a/hive-ag3nt/src/mcp.rs b/hive-ag3nt/src/mcp.rs
index fa87b57..14c2311 100644
--- a/hive-ag3nt/src/mcp.rs
+++ b/hive-ag3nt/src/mcp.rs
@@ -149,6 +149,9 @@ impl AgentServer {
     async fn send(&self, Parameters(args): Parameters<SendArgs>) -> String {
         let log = format!("{args:?}");
         let to = args.to.clone();
+        if let Err(refusal) = check_send_allowed(&to) {
+            return run_tool_envelope("send", log, async move { refusal }).await;
+        }
         run_tool_envelope("send", log, async move {
             let resp = client::request::<_, hive_sh4re::AgentResponse>(
                 &self.socket,
@@ -627,6 +630,54 @@ pub fn builtin_tools_arg() -> String {
 /// `mcp__<key>__<tool>` pattern in `--allowedTools`.
 const EXTRA_MCP_PATH: &str = "/etc/hyperhive/extra-mcp.json";
 
+/// Where the NixOS module writes the per-agent send allow-list (see
+/// `nix/templates/harness-base.nix`). Empty list = unrestricted (the
+/// default). Non-empty list constrains `mcp__hyperhive__send`'s `to`
+/// field; the manager is always implicitly permitted regardless of
+/// the list contents.
+const SEND_ALLOW_PATH: &str = "/etc/hyperhive/send-allow.json";
+
+/// Enforce the per-agent send allow-list. Returns `Ok` when the
+/// recipient is permitted (no list configured, manager always
+/// allowed, or `to` is in the list); returns `Err(refusal)` with a
+/// claude-readable string when blocked — the harness surfaces the
+/// refusal as the tool result so claude knows the message didn't
+/// land and can react (e.g. route via the manager instead).
+fn check_send_allowed(to: &str) -> Result<(), String> {
+    if to == hive_sh4re::MANAGER_AGENT {
+        // Always allow agents to talk to the manager — otherwise a
+        // misconfigured allow-list could leave a sub-agent unable
+        // to ask for help.
+        return Ok(());
+    }
+    let Ok(raw) = std::fs::read_to_string(SEND_ALLOW_PATH) else {
+        return Ok(()); // file missing → no policy configured → unrestricted
+    };
+    let allow: Vec<String> = match serde_json::from_str(&raw) {
+        Ok(v) => v,
+        Err(e) => {
+            tracing::warn!(
+                path = SEND_ALLOW_PATH,
+                error = ?e,
+                "send allow-list parse failed; falling back to unrestricted",
+            );
+            return Ok(());
+        }
+    };
+    if allow.is_empty() {
+        return Ok(()); // empty list = unrestricted (back-compat)
+    }
+    if allow.iter().any(|n| n == to) {
+        return Ok(());
+    }
+    Err(format!(
+        "send refused: recipient '{to}' not in hyperhive.allowedRecipients \
+         (configured in agent.nix). Allowed: {allow:?}. The manager is \
+         always reachable — route through `send(to: \"manager\", …)` if \
+         you need to reach someone outside the allow-list."
+    ))
+}
+
 #[derive(Debug, serde::Deserialize)]
 struct ExtraMcpServer {
     command: String,
diff --git a/nix/templates/harness-base.nix b/nix/templates/harness-base.nix
index 8aa6c31..2e682fa 100644
--- a/nix/templates/harness-base.nix
+++ b/nix/templates/harness-base.nix
@@ -5,6 +5,29 @@
   # this. The systemd service that actually runs the harness binary
   # differs per role and lives in the child module.
 
+  options.hyperhive.allowedRecipients = lib.mkOption {
+    type = lib.types.listOf lib.types.str;
+    default = [ ];
+    example = [ "alice" "manager" ];
+    description = ''
+      Names this agent is allowed to `send` to via
+      `mcp__hyperhive__send`. Empty list (the default) means
+      unrestricted — the agent can message any peer, the
+      operator, or the manager. Non-empty list constrains the
+      surface: only the listed names + the manager (always
+      allowed) get through; anything else returns an error
+      string to claude without touching the broker. The
+      operator (`operator`) needs to be in the list if the
+      agent should be able to surface output on the
+      dashboard.
+
+      Useful for sandboxing untrusted sub-agents — set
+      `[ "manager" ]` to scope them to manager-only chatter.
+      The manager itself is always exempt; this option only
+      affects sub-agent `send`.
+    '';
+  };
+
   options.hyperhive.extraMcpServers = lib.mkOption {
     type = lib.types.attrsOf (lib.types.submodule {
       options = {
@@ -63,6 +86,9 @@
     environment.etc."hyperhive/extra-mcp.json".text =
       builtins.toJSON config.hyperhive.extraMcpServers;
 
+    environment.etc."hyperhive/send-allow.json".text =
+      builtins.toJSON config.hyperhive.allowedRecipients;
+
     boot.isNspawnContainer = true;
 
     # `claude-code` is unfree. Each per-agent container's nixosConfiguration