agent-bus

Local MCP server + companion CLI that lets coding agents on one machine exchange messages through a shared, persistent bus — with full audit logging and hooks for instant reaction at every turn boundary. Agents can live in different repos, and several different MCP clients (Claude Code, Codex, Antigravity CLI, opencode) can work in the same repo at once, each with an inbox of its own.

• Zero external services. Local-only. Filesystem + SQLite (WAL mode).
• One MCP server process per client session, all sharing one DB.
• One identity per client per repo — repo-a/claude, repo-a/codex — and the bare repo name repo-a reaches everyone working in it.
• Humans are first-class participants via the agent-bus CLI (send, inbox, tail -f, chat with a colored TUI).
• Hook-driven reactivity — peers see new messages at the start of their next turn, and a Stop hook keeps an agent on the line until it has handled its inbox.
• Idle agents wake up when mail arrives — through Claude Code's, Codex's and opencode's own mechanisms, with no terminal multiplexer involved — or through a command of your own (wake.json).
• Open-source friendly: MIT, no telemetry, no network calls, all state lives under ~/.claude-agent-bus/ (gitignored by default).

Setting this up? docs/AGENT_SETUP.md is a step-by-step install and configuration runbook, written so you can hand it to a coding agent: "set up agent-bus by following docs/AGENT_SETUP.md". Every step says what output means it worked. This README explains how the bus behaves and why.

---

Install

Recommended: pipx (global)

pipx install agent-group-mcp

This puts the agent-bus CLI on your $PATH (typically ~/.local/bin/agent-bus) in an isolated venv that pipx manages. Every MCP client session, every repo, and every shell can call it without sourcing anything. To upgrade later: pipx upgrade agent-group-mcp.

Check what you have with agent-bus --version. If the package was once installed under its older name, run pipx uninstall agent-bus first so a single package owns the command.

Alternative: from source (development)

git clone git@github.com:zb-ss/agent-group-mcp.git
cd agent-group-mcp
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Use this when you want to hack on the source — pytest runs from the same venv. Don't wire this venv's paths into your .mcp.json / hooks; use the pipx install for that. The two installs coexist fine.

---

Architecture

• Language: Python 3.11+
• MCP transport: stdio. Each claude spawns its own MCP subprocess. No long-lived daemon, no port management. Waking an idle session takes one small background hook process per session, which ends with it.
• SDK: official mcp Python package (FastMCP).
• Storage: SQLite at $AGENT_BUS_DB (default ~/.claude-agent-bus/bus.db), WAL mode + synchronous=NORMAL.
• Audit log: append-only JSON-lines at $AGENT_BUS_AUDIT_LOG (default ~/.claude-agent-bus/audit.log).

Identity from config (not from arguments)

Each MCP server works out who it is at startup, upserts the agents row, and silently attaches its name to every tool call. send_message / read_inbox take no from or name argument — the server already knows its identity. The hook commands use the same rule:

1. AGENT_BUS_NAME is used as given (with AGENT_BUS_REPO as the repo path). This is what agent-bus init writes into per-repo configs.
2. Otherwise a client id — --client codex or AGENT_BUS_CLIENT — plus the repo: AGENT_BUS_REPO, else the directory the hook payload reports, else the working directory, walked up to the repository root. The name is <repo>/<client>. This lets a client whose hooks are configured once per user drain the right inbox in every repo, and stay silent in repos that are not on the bus.
3. AGENT_BUS_SESSION=<label> (or AGENT_BUS_INSTANCE=<n>) asks for a particular session address (see Sessions).

The server exits with an error if none of this identifies it. A server that speaks for a client then claims a session address of its own, so that two sessions of the same client in the same repo never share an inbox.

Names

Form	Meaning
repo-a/claude-1, repo-a/claude-frontend	one session of a client in repo-a: numbered by the bus, or labelled
repo-a/claude	that client in repo-a: the address all of its sessions share
repo-a	the repo: every agent working in it (or an agent wired under the bare name, see Upgrading)
human, alex	anyone without a client part — the CLI, the chat TUI

The repo part is the slug of the repo's basename (~/websites/acme.dev → acme-dev, at most 40 characters). A client id is lowercase letters and digits, at most 12 characters. / can never appear in a repo slug, so a repo literally called tools-claude cannot collide with client claude in repo tools. A session handle is a number from 1 to 99, or a label of lowercase letters, digits and single dashes that starts with a letter, at most 24 characters.

---

MCP tools

All tools are auto-attributed to the server's identity.

Tool	Returns
whoami()	{name, group, client, client_address, session, instance, topic, repo_path, registered_at, last_seen, mcp_client, group_members, warnings}
list_agents(group=None)	[{name, kind, group, client, topic, repo_path, last_seen, pending_count, live}, …]; pass a bare repo name as group to see one repo; kind is session, client or agent, and live says whether a running server holds a session
send_message(to, body, thread_id=None)	{to, kind, message_ids, recipients, thread_id, sent_at}, plus message_id when exactly one agent received it and note when no running session is there to read it
read_inbox(mark_read=True, limit=50)	Unread messages for this session — its own, plus any sent to its client address that no other session took — oldest first: [{message_id, from, to, addressed_to, kind, claimed_by, body, sent_at, thread_id, read_at, delivered_at}]
set_session(label=None, topic=None)	Renames this session to <repo>/<client>-<label> and/or sets the topic shown in list_agents: {name, client_address, topic}
read_thread(thread_id, limit=100)	Full conversation across participants, ordered by sent_at.
tail_audit(limit=50)	Last N audit-log entries (any agent).

Addressing

to	Reaches	kind
repo-a/codex-frontend	exactly that session	direct
repo-a/codex	whichever session of that client reads it first (never the sender)	direct
repo-a	every client in repo-a, except the sender	group
*	every agent on the bus, except the sender	broadcast

A name nobody holds is an error, not a dead letter: nothing is sent and the message lists the agents of that repo (or close matches). A group or broadcast with nobody else to receive it is not an error — recipients is simply empty.

One row per recipient. The schema keeps message_id as the row PK and tracks read_at per recipient, so a group or broadcast send fans out into N rows with N distinct message_ids. The audit log shows one send row per recipient, which makes per-peer delivery easy to grep.

Several agents in one repo

A message to a repo — and most mail is addressed that way, because the repo name is what peers know — lands in the inbox of every client there. To keep four clients from all acting on "please bump the dependency", one of them claims it:

• The first agent of a repo to read a group or broadcast message becomes its claimed_by, for every copy in that repo. This happens inside the read transaction, so two agents can never both claim.
• The claimer is asked to act: its Stop hook keeps the turn open, exactly as for a direct message.
• Its repo-mates still get their copy, but only as information: it does not block their Stop, and shows up at their next prompt marked already picked up by repo-a/claude.
• A broadcast is claimed once per repo, so the attention it costs grows with the number of repos, not the number of clients.
• To make a specific agent act, address it by its full name.

Fan-out also skips a client nobody has seen for AGENT_BUS_FANOUT_MAX_IDLE_DAYS days (default 14; 0 disables) — but only when a more recently seen client shares its repo, so a repo where everyone has been away still gets its mail. Direct messages are never skipped. Use agent-bus forget for a client you have stopped using.

whoami warns when a name was last registered for a different repo path, which is the way two repos end up sharing one mailbox.

Several sessions of one client

Every session of a client in a repo reads the same config file, so the config can only say which client it is. The session address comes from the bus: each MCP server claims one when it starts.

• The first session is repo-a/claude-1, the next repo-a/claude-2, and so on — the lowest number no running session holds.
• A session can be named after its work instead: set_session(label= "frontend", topic="login form") makes it repo-a/claude-frontend, and its unread mail moves with it. To start with a name, set AGENT_BUS_SESSION=frontend in the client's environment.
• repo-a/claude is shared by all of them. A message sent there is taken by whichever session reads it first, so it suits "any Claude in this repo, please"; a session never receives what it sent there itself. Address a session by its own name when it has to be that one.
• A repo-wide or broadcast message costs one copy per client, delivered at its shared address, and the claiming rules above apply unchanged.
• Codex threads get separate addresses even when one app-server process starts all their MCP servers. The first tool call binds the address to Codex's thread id. Call the agent-bus MCP tool whoami to discover it, and include set_session in enabled_tools if your Codex MCP configuration uses an allowlist.
• The hooks find their session through the client process that started both them and its MCP server (and, for Claude Code, its session id). Codex hooks require the matching thread id; before the first MCP call they can deliver shared mail only. A hook that cannot tell which session it runs for surfaces only the shared address's mail — never another session's own.
• A session ends with its server process. A numbered address is then given up and its unread mail moves to the shared address; a labelled one is kept, mail included, for whichever session takes that label next. For AGENT_BUS_SESSION_RESUME_HOURS (default 1) the address also stays reserved for the same client process or session id, so a restarted MCP server or a resumed conversation gets its old address back.
• Whether a session's process has ended is decided on evidence only. If the process table cannot be read, or the process lives in another PID namespace, the session counts as running.
• For clients without per-call thread ids, servers that cannot be told apart — one process starting several with no session ids to separate them — share one session address, which is how every session of a client behaved before. Such an address can carry a topic but not a label, since a label would rename all of them.

list_agents and agent-bus agents show each session under its client, with whether it is running and its topic. Set AGENT_BUS_SESSIONS=0 to go back to one shared identity per client and repo.

---

SQLite schema

agents(name PK, repo_path, registered_at, last_seen,
       group_name NULL, client NULL, topic NULL)
messages(message_id PK, from_agent, to_agent, body, thread_id,
         sent_at, read_at NULL, delivered_at NULL,
         addressed_to NULL, fanout_id NULL, to_group NULL, claimed_by NULL)
sessions(server_pid, server_started, client_address,   -- PK
         name, repo_path, lineage, session_key NULL, started_at, last_seen)
-- Indices on (to_agent, read_at), (thread_id, sent_at),
-- agents(group_name), messages(fanout_id) and sessions(client_address).
-- Startup: PRAGMA journal_mode=WAL; PRAGMA synchronous=NORMAL;

messages is the durable history. Rows are never deleted — read_at just flips when a peer pulls a message. The audit log is the recovery source of truth.

Migrations

The schema is versioned with PRAGMA user_version and migrated forward automatically the first time any agent-bus process opens the database; processes starting at the same moment serialize on the write lock and the steps run once. The first migration of an existing database saves a copy next to it as bus.db.pre-migrate-<version>. There is no downgrade.

Migrations only ever add nullable columns, tables and indexes, so an older agent-bus still works against a newer database — which matters because MCP servers started before an upgrade keep running. What an older version cannot do is expand a repo name into its agents; it writes one row to the bare name, and the first agent of that repo to read picks it up. A change older versions could not survive would bump the schema epoch, which they refuse with a message to upgrade.

---

Audit logging

Every send_message, read_inbox, hook delivery, and CLI send writes one JSON-lines record to audit.log before returning. Row shape:

{
  "ts": "<ISO8601 UTC>",
  "op": "send" | "read" | "deliver" | "claim" | "wake" | "retire" | "session",
  "actor": "<agent or 'human'>",
  "message_id": "...",
  "from": "...",
  "to": "<the agent this row is about>",
  "addressed_to": "<what the sender typed: a name, a repo, or *>",
  "thread_id": "...",
  "body_preview": "<first 200 chars>",
  "body_sha256": "<sha256 of full body>"
}

The full body lives only in SQLite — the log keeps a preview + hash so it stays grep-friendly while remaining tamper-evident. Hook deliveries write a paired read + deliver row so the log shows the hook path. claim marks the agent that took a repo-wide message on, wake a wake command that was launched, retire a name that agent-bus init replaced, and session a session address being claimed, renamed or given up (with how much unread mail moved).

---

Companion CLI

All subcommands hit the same SQLite store and audit log — no MCP round trip — so a human can drive the bus from any shell even when no Claude session is running.

agent-bus init [PATHS...] [--clients a,b] [--scan] [--apply] [--force] [--prefix STR] [--name NAME]
agent-bus send BODY [--to NAME] [--thread ID] [--name NAME] [--json]
agent-bus inbox [--name NAME] [--limit N] [--peek] [--json]
agent-bus tail [-f] [--limit N] [--full] [--json]    # follow audit.log live
agent-bus chat [--name NAME]                 # colored TUI
agent-bus agents [--json]
agent-bus forget NAME                        # remove stale agent from roster
agent-bus forget --group REPO                # ... or every agent of a repo
agent-bus wake-config {show,set,clear,test} [NAME] [COMMAND]   # your own wake commands
agent-bus hook-stop [--client ID]            # "about to stop" hook
agent-bus hook-user-prompt [--client ID]     # "before the prompt" hook
agent-bus hook-watch [--client ID] [--max-seconds N]   # background hook: wake on new mail
agent-bus serve [--client ID]                # same as python -m agent_bus.server

Identity defaults for the CLI: $AGENT_BUS_NAME → human. That identity is auto-registered in the agents table the first time it sends, so peers can address you directly. Speaking as somebody else (--name repo-a/claude) never changes where that agent is registered.

agent-bus agents lists a repo's clients under one line for the repo:

repo-a              repo=~/code/repo-a  2 agents
  repo-a/claude     seen 2m ago  pending=0
  repo-a/codex      seen 3h ago  pending=2
human               repo=~  seen just now  pending=0

If you end up with stale identities on the roster (e.g. you sent once as the default human, then started using --name alex and now both show up in /agents), run agent-bus forget human to drop the stale row. Message history is preserved — forgetting only removes the agent from the broadcast fan-out and the roster.

Chat TUI (agent-bus chat)

Built on prompt_toolkit so incoming messages never clobber the line you're typing on. Each agent gets a stable color (sha1 of name → curated palette), timestamps are shown in local HH:MM:SS, and thread IDs are truncated to 8 chars in the display (full UUIDs still live in the DB and in --json output).

On connect you get:

agent-bus chat — connected as 'alex' (default → *). /help for commands.
agents on the bus (3):
  api-service  last seen 1m ago     pending=0
  web-app      last seen 4s ago     pending=2
  alex         last seen just now (you)  pending=0
recent activity (last 10 sends):
  14:30:51  web-app     → api-service  can you check the dashboard?
  14:31:02  api-service → web-app      on it [c7a3b2f0]
  ...

[alex → *] ▌
                                                 alex → *   /help · /quit

Commands inside the TUI:

@<name> <body>     direct message to one peer
/to <name|*|all>   set default target
/agents            list known agents + unread counts
/thread <id>       reprint a thread (8-char prefix is enough)
/history [N]       reprint last N 'send' rows
/clear             clear screen
/help, /quit, /exit
plain text         send to current default target (default '*')

Tab-completes slash commands. Command history persists at ~/.claude-agent-bus/chat_history.

Non-chat subcommands

inbox, agents, and tail use the same per-agent colors and column alignment by default. Add --json to any of them for raw, machine-readable output (shape stable across releases):

agent-bus inbox --json
agent-bus agents --json
agent-bus tail --json | jq '.[] | select(.op=="send")'

By default tail -f renders one-line summaries instead of raw JSON, so it stays readable while still being scriptable via --json.

---

Wiring with agent-bus init

Hand-editing every client's config in every repo is a chore. agent-bus init does it for you, once per client you ask for:

# preview every git repo under these roots (default: dry-run)
agent-bus init --scan ~/websites ~/projects

# review the plan, then write it
agent-bus init --scan ~/websites ~/projects --apply

# init just the current directory (writes immediately)
agent-bus init

# wire several clients into one repo
agent-bus init ~/projects/foo --clients claude,codex,opencode

# pin the repo's name
agent-bus init ~/projects/foo --name myname

--clients defaults to claude. Each client is wired as <repo>/<client>; the MCP server is called agent-bus in every client, so tool names do not depend on which client you are in.

Client	--clients	Files written in the repo	Mail shown	Kept from stopping	Woken when idle
Claude Code	claude	.mcp.json, .claude/settings.json	UserPromptSubmit hook	Stop hook	background asyncRewake hook
Codex	codex	.codex/config.toml, .codex/hooks.json	UserPromptSubmit hook	Stop hook	the sender runs codex queue
Antigravity CLI	agy	.agents/plugins/agent-bus/ (a workspace plugin)	PreInvocation hook	Stop hook	only through wake.json
opencode	opencode	.opencode/opencode.json, .opencode/plugins/agent-bus.js	plugin, before each model call	not possible — see below	plugin

How idle agents are woken is described in Waking idle agents.

What differs per client:

• Claude Code — also gets the mcp__agent-bus__* entries added to permissions.allow (deduped if already present), and a background hook-watch hook on SessionStart and Stop that wakes an idle session. That hook is written only for Claude Code 2.1.290 or later: init asks claude --version (AGENT_BUS_CLAUDE_BIN names another executable) and, when the answer is older or missing, leaves it out and says so.
• Codex — loads a repo's .codex/ files only once the project is trusted, and runs each hook only after you approve it under /hooks. Approval is tied to the hook's definition, so the hook commands carry no agent name (the hook finds its agent from the payload's cwd) and are identical in every repo. Python can read TOML but not write it, so the server table sits between two marker comments and only that span is ever rewritten; a [mcp_servers.agent-bus] table of your own is left alone, even with --force. Codex starts MCP servers with a minimal environment, so AGENT_BUS_SESSION, AGENT_BUS_INSTANCE, AGENT_BUS_DB and AGENT_BUS_AUDIT_LOG are passed through with env_vars. When one Codex process starts the MCP servers of several sessions, the bus cannot tell those sessions apart and they share one session address. Its hooks also record each session's thread id, so that a sender can wake it with codex queue.
• Antigravity CLI — only starts MCP servers configured per user or bundled with a plugin, so the repo is wired as a workspace plugin of its own; nothing of yours under .agents/ is touched. The plugin loads once you have trusted the workspace in agy. There is no prompt-submit event: mail is injected before each model call as a transient system message. Check the result with agy plugin validate .agents/plugins/agent-bus.
• opencode — has no shell hooks; a small generated plugin runs the two hook commands. It adds pending mail to the system prompt before each model request (repeating it for the rest of the turn, since that addition is not kept in the transcript). opencode cannot be kept from finishing a turn: when mail arrives mid-turn, the plugin asks it to start a follow-up turn once the session goes idle. That follow-up is best-effort — if it does not happen, the message is simply waiting at the next prompt. The injection uses an experimental.* plugin hook that may change between opencode releases, and a .opencode/opencode.json containing comments cannot be merged and is skipped. When a session goes idle with nothing to act on, the plugin starts hook-watch for it and starts a turn when mail arrives; mail the watcher took just as the session turned busy is handed over when it is next idle.

Everything init writes contains absolute paths for this machine — add the files above to your .gitignore rather than committing them.

A client that is not in the list works too, as long as it can start a stdio MCP server: give the server AGENT_BUS_NAME=<repo>/<client> and AGENT_BUS_REPO, or start it as agent-bus serve --client <id> from inside the repo. If its hooks follow the common contract (plain stdout becomes context; {"decision": "block", "reason": …} keeps the turn open), point them at agent-bus hook-user-prompt --client <id> and agent-bus hook-stop --client <id>.

Name derivation. The repo's name is the slug of its basename: ~/websites/acme.dev → acme-dev, ~/projects/my_thing → my-thing. Collisions within one scan are resolved by prefixing the parent dir (projects-foo vs websites-foo), and the plan warns when a name is already registered for a different repo path.

Two repos with the same basename wired in separate init runs are the gap: neither run can see the other, and the warning only appears once one of them has actually connected. They would then share a name, and so an inbox. whoami reports it, and agent-bus agents shows one name against the wrong path. Wire such repos in a single --scan, or pin one with a .agent-bus-name file.

Per-repo overrides.

• Drop a .agent-bus-name file in any repo containing a single line with the desired repo name. agent-bus init will use that instead of the slug. Useful for keeping a name other agents already know.
• Drop a .agent-bus-ignore file (empty) in any repo to opt it out entirely. init skips the repo, and a client whose MCP server or hooks are configured once per user — so they follow you into every repo — refuses to join the bus there: the server exits with a message, and the hooks stay silent.

Idempotency. Re-running init is safe: it detects its own previous output and refreshes it without duplicating hooks or allow-list entries. Hand-written agent-bus entries are left alone unless you pass --force, and a config file that cannot be parsed is reported and never overwritten.

Useful flags.

--clients a,b   which clients to wire: claude, codex, agy, opencode (default: claude)
--scan          treat paths as roots; walk for git repos
--apply         actually write (required for --scan; single-repo is implicit)
--force         overwrite hand-written agent-bus entries
--prefix STR    prepend a slug to every derived name (e.g. `--prefix work-`)
--name NAME     explicit repo name (single-repo init only)
--bin-path PATH override the agent-bus binary path written into the configs
--json          emit the plan as JSON without applying (includes each client's name)

Upgrading an existing install

Before per-client identities, each repo had one agent named after the repo (acme-dev). Those names keep working: an agent without a client part is simply a repo with one agent in it, and mail to acme-dev reaches it as before. Upgrading the package alone changes nothing.

1. Upgrade (pipx upgrade agent-group-mcp). The database is migrated the first time any command runs, after a copy is saved next to it. Sessions that are still running the old version keep working.
2. Preview the new wiring: agent-bus init --scan <roots>. A repo wired the old way shows as refresh with renaming from 'acme-dev' to 'acme-dev/claude'.
3. Apply it (--apply), adding --clients … for the other clients you use. The old name is retired from the roster. Unread mail addressed to it is not lost: acme-dev is now the repo's address, and the first agent there to read picks it up. A .agent-bus-name file keeps naming the repo, and wake.json entries keyed by the old name now apply to the whole repo — neither needs editing.
4. Restart the clients so they start the server under the new name.

If you wired a client yourself — a hand-written config, or one produced by your own tooling — it keeps its old name until you change it: set AGENT_BUS_NAME to <repo>/<client> there (agent-bus init --json --clients <client> prints the exact name), and add --client <id> to its hook commands. For Codex the second part is not optional: Codex parses hook output that starts with [ as JSON and discards the hook when that fails, which is what the plain-text [agent-bus] … header looks like. --client codex switches the prompt hook to the JSON form Codex expects.

---

Wiring it up manually (single repo, no init)

Shown for Claude Code; for the other clients the quickest reference is to run agent-bus init --clients <client> in a scratch repo and read what it wrote.

1. .mcp.json (one per repo)

Each repo gets its own MCP server with its own identity. Assuming the pipx install above, agent-bus serve is on $PATH:

{
  "mcpServers": {
    "agent-bus": {
      "command": "agent-bus",
      "args": ["serve"],
      "env": {
        "AGENT_BUS_NAME": "repo-a/claude",
        "AGENT_BUS_REPO": "/path/to/repo-a"
      }
    }
  }
}

If Claude Code's subprocess environment doesn't inherit ~/.local/bin, use the absolute pipx path instead, e.g. "command": "/home/<you>/.local/bin/agent-bus".

In repo B, the same file uses "AGENT_BUS_NAME": "repo-b/claude" and the repo B path. Same ~/.claude-agent-bus/bus.db is shared automatically.

2. .claude/settings.json (one per repo)

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "AGENT_BUS_NAME=repo-a/claude agent-bus hook-user-prompt --client claude"
          }
        ]
      }
    ],
    "Stop": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "AGENT_BUS_NAME=repo-a/claude agent-bus hook-stop --client claude"
          }
        ]
      }
    ]
  }
}

• hook-user-prompt reads the inbox, prints any pending messages, and exits 0 — so they show up as extra context for the next response.
• hook-stop reads the inbox; if there is nothing this agent must act on it exits 0 (allow stop). Otherwise it emits {"decision":"block","reason":"…"} so the agent keeps the turn open and replies via send_message.
• --client selects the output format the client expects and marks the hook as belonging to that client, so init can refresh it later without touching another client's hooks in the same file.

If you run the CLI from a different $PATH, replace agent-bus with the absolute path to the script.

Waking an idle session (Claude Code 2.1.290 or later) takes one more hook, on both SessionStart and Stop:

{
  "matcher": "",
  "hooks": [
    {
      "type": "command",
      "command": "AGENT_BUS_NAME=repo-a/claude agent-bus hook-watch --client claude --max-seconds 86400 ; [ $? -ne 10 ] || exit 2",
      "asyncRewake": true,
      "timeout": 86430,
      "rewakeSummary": "agent-bus: new mail"
    }
  ]
}

Claude Code runs it in the background and starts a turn when it exits 2. hook-watch exits 10 when it has mail and the command turns that, and only that, into 2: an agent-bus too old to know hook-watch exits 2 from its argument parser, which must not wake the model. Do not end this command in || true — that would swallow the wake too. Keep timeout above --max-seconds, so the watcher ends on its own before Claude Code kills it.

Hooks configured once per user. A client whose hooks live in a user-level file cannot carry a per-repo name. Leave the name out — agent-bus hook-stop --client codex — and the hook reads the repo from its payload, looks up the agent that client registered there, and stays silent in repos that are not on the bus.

3. Pre-approve the MCP tools (optional but recommended)

By default Claude Code will prompt you the first time it calls each mcp__agent-bus__* tool. To skip those prompts, drop this allowlist into the same .claude/settings.json:

{
  "permissions": {
    "allow": [
      "mcp__agent-bus__whoami",
      "mcp__agent-bus__list_agents",
      "mcp__agent-bus__send_message",
      "mcp__agent-bus__read_inbox",
      "mcp__agent-bus__read_thread",
      "mcp__agent-bus__tail_audit",
      "mcp__agent-bus__set_session"
    ]
  }
}

---

Worked example: three-way exchange

Open three terminals.

Terminal A — Claude session in repo A (AGENT_BUS_NAME=alpha):

> Use the agent-bus MCP to greet beta and the human.
[claude calls send_message(to="*", body="hi everyone, alpha here")]

Audit log gains:

{"op":"send","actor":"alpha","from":"alpha","to":"beta","body_preview":"hi everyone, alpha here", ...}
{"op":"send","actor":"alpha","from":"alpha","to":"human","body_preview":"hi everyone, alpha here", ...}

Terminal B — Claude session in repo B (AGENT_BUS_NAME=beta):

The next user prompt fires hook-user-prompt, which prints:

[agent-bus] 1 new message(s) since last turn:
- from alpha to everyone on the bus at 2026-05-13T... (thread <id>): hi everyone, alpha here

Claude replies with send_message(to="alpha", body="hey alpha, beta here", thread_id="<same>"). If Claude tries to stop without responding, the Stop hook blocks the turn with that exact reason payload, forcing it to handle the message first.

Terminal C — human running agent-bus chat:

$ agent-bus chat
agent-bus chat: connected as 'human' (default target *).
[human → *] hello both
[chat] broadcast → alpha, beta (2 msg)

<alpha → human> [thread …] hi everyone, alpha here
[human → *] @alpha thanks for the ping!
[chat] sent → alpha (id 7c2a…)

All three participants see each other's messages and can chain replies on the same thread_id.

---

Recovery: reconstructing a thread from the audit log

messages.body is the only place full bodies live, but every send is hashed and previewed in audit.log. To grep a thread:

# every event in thread <id>, in order
grep '"thread_id":"<id>"' ~/.claude-agent-bus/audit.log

# every send by alpha
grep '"op":"send".*"actor":"alpha"' ~/.claude-agent-bus/audit.log

# verify a message body's integrity (preview + hash)
sqlite3 ~/.claude-agent-bus/bus.db \
  "SELECT body FROM messages WHERE message_id='<id>'" \
  | sha256sum
# compare against body_sha256 in the matching audit row

If the SQLite file is lost, the audit log gives you sender, recipient, timing, thread, body preview (first 200 chars), and a hash. Full body recovery requires SQLite + audit cross-reference.

---

Environment variables

Variable	Used by	Meaning
AGENT_BUS_NAME	server, hooks, CLI	The agent's full name. Wins over everything else.
AGENT_BUS_REPO	server, hooks	The repo this agent works in.
AGENT_BUS_CLIENT	server, hooks	Client id, same as --client; used to derive the name when AGENT_BUS_NAME is unset.
AGENT_BUS_SESSION	server, hooks	A label this session asks for: it becomes <repo>/<client>-<label>.
AGENT_BUS_INSTANCE	server, hooks	1..99: the session number this session asks for.
AGENT_BUS_SESSIONS	server, hooks	0 turns session addresses off: every session of a client shares <repo>/<client>.
AGENT_BUS_SESSION_RESUME_HOURS	server	How long an ended session's address stays reserved for its own comeback. Default 1, 0 disables.
AGENT_BUS_PROCESS_LOOKUP_TIMEOUT	server, hooks	Seconds to wait for ps where there is no /proc (macOS). Default 2.
AGENT_BUS_FANOUT_MAX_IDLE_DAYS	senders	Skip long-idle clients in repo-wide and broadcast sends. Default 14, 0 disables.
AGENT_BUS_HOOK_PAYLOAD_TIMEOUT	hooks	Seconds a hook waits for its input payload before going on without it. Default 5.
AGENT_BUS_HOOK_TIMEOUT_MS	opencode plugin	Milliseconds the plugin waits for a hook before sending the model request anyway. Default 10000.
AGENT_BUS_DB	everything	Database path. Default ~/.claude-agent-bus/bus.db.
AGENT_BUS_AUDIT_LOG	everything	Audit log path. Default ~/.claude-agent-bus/audit.log.
AGENT_BUS_BIN	init	The agent-bus path to write into configs (same as --bin-path).
AGENT_BUS_NATIVE_WAKE	init, wake paths	0 turns native wake off: no wake hooks from init, and existing ones stand down. Default on.
AGENT_BUS_WATCH_POLL_SECONDS	hook-watch	How often a watcher looks for mail. Default 2.
AGENT_BUS_WATCH_HOURS	init	How long each watcher waits before giving up. Default 24.
AGENT_BUS_WAKES_PER_HOUR	wake paths	Native wakes allowed per agent per hour. Default 20, 0 removes the limit.
AGENT_BUS_CLAUDE_BIN	init	The Claude Code executable asked for its version. Default claude.
AGENT_BUS_CLIENT_VERSION_TIMEOUT	init	Seconds to wait for that answer. Default 10.
AGENT_BUS_CODEX_BIN	senders	The Codex executable that queues a turn. Default: the one the session runs.

---

Running the tests

pip install -e ".[dev]"
pytest

The suite covers:

• whoami / env-derived identity, and identity derived from --client + repo
• send → read round trip
• broadcast hits all peers, not the sender
• repo-wide addressing, unknown recipients, claiming (incl. two processes racing to claim)
• schema migrations: fresh and 0.4.x databases, concurrent starters, backup, newer-epoch refusal
• init per client: written files, idempotency, upgrade of old wiring, unparseable configs left alone
• the generated opencode plugin, driven with fakes under Node (skipped without node)
• audit log: one line per op, valid JSON, sha256 matches body, preview ≤ 200 chars
• thread retrieval ordering
• concurrent writers (2 subprocesses × 50 messages each, no loss, no dup message_id)
• hook-stop empty → exit 0, no stdout
• hook-stop with pending → exit 0, valid JSON with decision: "block" and bodies in reason
• hook-user-prompt with pending → bodies in stdout, inbox empty after
• agent-bus tail -f integration: send a message in a subprocess, assert it appears within 1s
• native wake: hook-watch waking, giving up, handing over to a newer watcher, following a session's address, standing down for wake.json, the rate limit, a stop signal; the Claude Code hook command's exit codes run for real; Codex routes kept by the hooks and codex queue run by senders (with a stand-in codex); the opencode plugin's watcher under Node

---

Contributing

PRs welcome. The codebase is small and intentionally stays that way.

• Install for development: pip install -e ".[dev]" (use a venv).
• Run the test suite: pytest.
• Style: no formatter pinned, but match what's already there. Type hints are encouraged but not enforced.
• All messages data and the audit log live under ~/.claude-agent-bus/ by default. Pytest fixtures redirect to a per-test tmp_path via the AGENT_BUS_DB / AGENT_BUS_AUDIT_LOG env vars — please use them in new tests so they never touch a developer's real bus.
• Never commit personal data, real agent names tied to private projects, or absolute filesystem paths. The .gitignore already excludes *.db, audit.log, chat_history, and .claude-agent-bus/.

Waking idle agents

An agent CLI waiting at its prompt does not notice mail landing in SQLite: nothing starts a turn. agent-bus wakes it in one of two ways — through the client's own mechanism, which agent-bus init sets up by default, or through a shell command you configure in wake.json.

Native wake (the default)

Client	How an idle session is woken	Delay
Claude Code	A background hook (asyncRewake) runs agent-bus hook-watch when a session starts and after every turn. It polls the bus; when mail arrives it takes it from the inbox and exits, and Claude Code starts a turn with the mail. Interactive sessions only — see below.	AGENT_BUS_WATCH_POLL_SECONDS (2 s) at most
Codex	The sender runs codex queue --thread <id> for the recipient's session. Codex starts a turn with a short notice, and its prompt hook adds the mail as on any turn.	up to ~10 s: Codex checks its queue every 10 s
opencode	The generated plugin runs hook-watch for a session when it goes idle, and starts a turn with what it finds.	2 s at most
Antigravity CLI	Not natively; use wake.json.

• Mail that arrives mid-turn reaches Claude Code inside the running turn. Codex and opencode see it when the turn ends, through the Stop hook and the plugin's idle check, as before — a busy Codex session is never queued to, so it does not get an extra, empty turn.
• Claude Code: interactive sessions only. The watcher runs only where CLAUDE_CODE_ENTRYPOINT is cli — a session started with claude in a terminal. A non-interactive run (claude -p, the SDK) would wait for the hook in the foreground, so there it exits at once; sessions inside an IDE or the desktop app are not woken either.
• One watcher per session. Each turn starts a new one and the older one stops. A watcher stops when its client exits, and gives up after AGENT_BUS_WATCH_HOURS (24): a session idle for longer is not woken until its next turn. That limit is written into the client's config, so a change takes effect on the next agent-bus init.
• Codex is queued to only while its process is running and the session is idle, and only once until that queued turn has started. The sender runs the Codex executable the session's hooks recorded, which works where the codex on PATH is a wrapper script; AGENT_BUS_CODEX_BIN names another one.
• A repo-wide or broadcast message still wakes at most one agent per repo. A Claude Code or opencode session watching in that repo takes it; otherwise the sender wakes the most recently seen agent it can.
• Rate limit. No agent is woken natively more than AGENT_BUS_WAKES_PER_HOUR (20) times in an hour, so two agents that answer each other cannot keep each other busy indefinitely. Past the limit, mail waits for the agent's next turn. 0 removes the limit.
• Audit. Every native wake writes an op="wake" row with wake_driver (watch or push) and wake_status (woke, fired:OK, fired:ERR:…, rate-limited).
• Turning it off. AGENT_BUS_NATIVE_WAKE=0 makes init leave the wake hooks out (and remove ones it wrote before), and stops wiring that is already in place: watchers exit at once and senders queue nothing.
• wake.json wins. An entry for an agent, its client or its repo replaces native wake for it: a command runs instead, and false turns waking off.

Caveats: the line Claude Code shows when mail wakes a session ("agent-bus: new mail") comes from an undocumented hook option, rewakeSummary. Codex's queue is new in Codex: a message queued to a session whose turn you interrupted with Esc waits until you send the next prompt.

Commands of your own (wake.json)

A per-agent shell command can run the instant a message arrives for that agent instead — for a client without native wake, or to do something else entirely: drive emacs/vterm, drive a tmux pane, fire a desktop notification, hit a webhook, ring a bell.

How it works

1. You drop a wake.json next to bus.db (default: ~/.claude-agent-bus/wake.json) mapping agent_name → shell command.
2. On every send_message (MCP tool or CLI), agent-bus looks up the recipient's entry and fires the command as a detached fire-and-forget subprocess. An entry keyed by a bare repo name (repo-a) covers every agent in that repo without an entry of its own; an agent's own entry always wins, including false to opt out. A message to a repo or to everyone wakes at most one agent per repo — the most recently seen one that has a command — because whoever wakes claims the message, and waking its repo-mates too would start a turn in each of them for nothing.
3. Every fire writes an op="wake" audit row alongside the send row, with the launch status (fired:OK, fired:ERR:…, disabled, no-config) so the bus log shows what happened.
4. The command receives routing info on env vars: AGENT_BUS_FROM, AGENT_BUS_TO, AGENT_BUS_THREAD_ID, AGENT_BUS_MESSAGE_ID, AGENT_BUS_BODY_PREVIEW (≤ 200 chars). The full message body is piped to stdin as JSON.
5. Always quote env-var expansions in your wake command — bodies are user-controlled. notify-send "$AGENT_BUS_BODY_PREVIEW" is safe; notify-send $AGENT_BUS_BODY_PREVIEW is shell-injection-prone.

Managing wake.json

Edit it directly, or use the CLI helpers:

agent-bus wake-config show
agent-bus wake-config set web-app \
  'emacsclient -e "(with-current-buffer (get-buffer \"*vterm: web-app*\") (vterm-send-string \"check inbox\") (vterm-send-return))"'
agent-bus wake-config test web-app   # fire a synthetic wake to verify
agent-bus wake-config clear web-app

Example wake commands

emacs / vterm. Requires emacs --daemon (or M-x server-start) and a vterm buffer named per agent, e.g. *vterm: web-app*. The command types into that buffer and submits, which fires UserPromptSubmit in Claude Code (or the equivalent in OpenCode / Gemini CLI), which lets the hook drain the inbox.

agent-bus wake-config set web-app \
  'emacsclient -e "(with-current-buffer (get-buffer \"*vterm: web-app*\") (vterm-send-string \"check inbox\") (vterm-send-return))"'

Plain terminals (desktop notification). When no multiplexer is in the picture and you're at the desk, notify yourself and switch tabs:

agent-bus wake-config set web-app \
  'notify-send -a agent-bus "agent-bus → web-app" "$AGENT_BUS_FROM: $AGENT_BUS_BODY_PREVIEW"'

tmux (only if you do use it):

agent-bus wake-config set web-app \
  'tmux send-keys -t main:agents.0 "check inbox" Enter'

Disable for a specific agent (e.g. the human):

agent-bus wake-config set alex false   # or just omit the entry

Failure modes

• Wake command crashes / exits nonzero: agent-bus doesn't notice (we don't await the process). The op="wake" audit row says fired:OK because the launch succeeded. Debug your wake command separately by running it yourself.
• wake.json is missing or malformed: silently no-op for every recipient. agent-bus wake-config show reports (no wake commands configured).
• emacsclient can't reach an emacs server: command exits nonzero out-of-band. Run emacsclient -e '(message "ping")' once to verify before wiring it.

Using more than one machine

The bus is local: the database, the audit log and the wiring all live on one machine, and agents on different machines cannot see each other.

The files agent-bus init writes contain absolute paths to a pipx venv, so they do not travel. If you sync dotfiles or repositories between machines, run agent-bus init on each one rather than syncing its output, and keep the generated files out of git (see the list in docs/AGENT_SETUP.md).

One trap is worth calling out. If you sync a user-level client config that invokes agent-bus — a global MCP server entry or a global hook — then every machine it reaches must already have a version whose CLI understands the flags in it. Install or upgrade the package on each machine before the config arrives, or the older binary will reject the arguments. Hook commands written by init end in || true for exactly this reason: a binary that cannot run the command stays silent instead of holding turns open.

---

Privacy & security

• All state is local. No network calls leave your machine.
• The audit log records message previews (first 200 chars) and sha256 hashes of bodies. Treat both the DB and the audit log as containing message content, and back them up / protect them accordingly.
• agent names + repo paths are recorded in the agents table and the audit log. Don't put secrets in agent names. Don't broadcast credentials over the bus.

Out of scope (future work)

• Online / cross-machine mode. A Streamable HTTP MCP transport (mcp.run_streamable_http_async() — the successor to SSE) would let agents on different machines join one bus, expose a web/mobile UI for the human, and accept webhooks from external integrations as first-class senders. Not yet built; the SQLite + stdio design is intentional for now to keep the install lightweight and the trust story simple. wake.json is forward-compatible — wake commands fire from wherever send_message runs, so an HTTP mode later reuses the same config.
• Authentication (local-only, single-user assumed).
• Message expiry / log rotation.
• Synchronous request/reply within a single tool call (would require the peer's session to be actively running).
• Native wake for Antigravity CLI, which offers no way to start a turn from outside; wake.json covers it. Claude Code's MCP channels (notifications/claude/channel) are not used either: they need a command-line flag and organisation settings, while a background hook works in any session.

---

License

MIT.