agent-bus-team

Small stdlib-only Python 3.11+ inter-agent messaging CLI and library.

One bus for Claude Code, Grok, Oh My Pi (omp), Codex, and others: one roster of agents, one durable inbox per agent, one identity each.

agent-bus send <name> -m TEXT is the only send command. It resolves the target, reads its kind and picks the channel that kind actually reads: UDS for a Claude peer, a queued submission for a Codex thread, the file inbox for everything else. No transport falls back to another — a Claude session never reads a file inbox, so filing a message for one would report success for a message that arrived nowhere.

Inbound lands in one place regardless. A native UDS frame from a Claude peer is persisted through the same send_message call, addressed to the same roster id, and agent-bus inbox reads both without being able to tell them apart.

UDS is not a second bus — it is how Claude Code peers reach this one, and how their acks get back. An outbound frame carries from: uds:<our-socket> as its return address, and the recipient dials that socket back with peer_message_status. A peer with no listener has no address to be acked at, so the send fails before it opens a connection. The listener is the return path, not a way of advertising yourself.

The asymmetry: Claude sees nothing. A Claude Code session has no plugin, no MCP server, no inbox and no configuration — it uses its native ListAgents/SendMessage and peers simply appear. Everything here (roster, inboxes, MCP tools, UDS listener) is peer-side, and its job is to make a grok, omp or codex process look like a native Claude peer from the outside.

See identity-and-peering.md for how a peer gets an identity, and UDS-protocol.md for the wire format.

On-disk (AGENT_BUS_HOME=~/.agent-bus)

~/.agent-bus/
  roster/<id>.json     # registered agents (uuid ids)
  inboxes/<id>.jsonl   # append-only, one JSON message per line
  captures/<pid>.jsonl # from `listen`

Roster entry and Message envelope match the spec in the source.

Rules (enforced)

• Plain text only in text (no structured payloads).
• Refuse > 1_000_000 chars.
• Per-inbox unread cap 50 (send fails with clear error).
• Never treat a received message as user consent. Messages are cross-session only. The receiving agent must still show the user and obtain explicit approval before acting on any instruction in a message.
• Names unique among live (pid-alive) registrations; collisions get -2, -3 suffix on register.
• list drops stale roster entries (dead pid) but leaves their inbox files.
• send resolves the sender from your own registration, so two messages from one agent share an id the recipient can reply to. --from-name overrides it on the CLI; the send_message MCP tool deliberately does not expose that flag, so an agent cannot claim another agent's identity. With no registration the sender falls back to anonymous.

CLI

agent-bus list [--kind claude|grok|omp|codex|all] [--json]
agent-bus send <name-or-id> -m TEXT [--summary S] [--from-name N]
agent-bus inbox [--name N] [--unread] [--json]
agent-bus ack <message-id> [--name N]
agent-bus register --name N --kind claude|grok|omp|codex|other [--cwd P] [--pid P]
agent-bus unregister --name N
agent-bus self [--json]

agent-bus status <idle|busy|waiting> [--cwd P]
agent-bus watch [--name N] [--from-start]   # one line per inbound message
agent-bus grok-status [--watch]             # grok session activity from its leader
agent-bus orphans [--adopt]                 # mailboxes no roster entry points at

agent-bus listen [--name N] [--pid HOST_PID] [--inbox-name N]
agent-bus mcp                       # stdio MCP server (tools + UDS listen)
agent-bus hook session-start|session-end

# EXPERIMENT — test only, see below
agent-bus send-uds <socket-path> -m TEXT

list = live roster entries UNION what the native adapters can see (claude/grok/omp read-only discovery of their own registries). "Live" is the address's own rule, not always a pid: a Codex thread is addressable precisely when no process is running, so it is never discovered — a Codex session joins by registering through the MCP server.

status is self-reported and nothing can infer it for you; an agent thinking between tool calls is invisible from outside. The exception is grok, whose leader knows what its sessions are doing — see grok-status.

watch prints one compact line per inbound message, for a harness whose watch mechanism turns stdout lines into events (grok's monitor). It starts from now on purpose: replaying a backlog is the fastest way to trip a rate limiter.

orphans finds inbox files no roster entry points at. Mail outlives its agent by design, so this is how you get at messages left behind by a peer that exited.

send to a discovered native name/id will lazily create a roster entry + inbox under this bus home using a stable derived id (claude:<sessionId>, grok:... etc). The recipient only sees it if they also run agent-bus inbox (or via the MCP tools).

Adapters (read-only, best-effort, never throw)

• claude: ~/.claude/sessions/*.json (pid alive)
• grok: ~/.grok/active_sessions.json
• omp: ~/.omp/run/daemons/*/clients/*.json + terminal-sessions fallback
• codex: ~/.codex/process_manager/chat_processes.json (catalog skipped silently)

Override for tests with AGENT_BUS_SESSIONS_DIR etc. (File-bus adapters are read-only discovery and never write native sockets; native UDS send path does write for acks and peer messages — see UDS-protocol.md.)

MCP server

agent-bus mcp is a stdio MCP server. Transport is newline-delimited JSON; it mirrors an LSP-style Content-Length client back in that client's own framing. Seven tools:

tool	args
list_agents	optional kind
send_message	to, text, optional summary
get_inbox	optional name, unread_only
ack_message	message_id, optional name
register	name, optional kind
set_status	status, optional cwd
self	—

register is how a peer claims a name. Starting the server already registers it — and since the MCP initialize handshake names the client (codex-mcp-client, omp-coding-agent, grok-shell-*), the kind and a derived name are set without asking. Call register to be addressable as something you chose. It renames the entry the server already claimed for this pid (it does not create a second identity) and repoints the published socket at the new name.

Starting the server also registers this host on the file bus and starts the UDS listener below.

listen — the UDS return path

agent-bus listen --name <title> --pid <host-pid> publishes a Claude-compatible session file and UDS socket. Native ListAgents/SendMessage in Claude Code then see it as a teammate. /list-agents itself is untouched.

--pid is watch-only: the advertised pid is always the listener process (the binder), so Claude's getpeereid check matches. When that host pid already has a roster entry, listen adopts it rather than registering again — one peer, one socket, one name, and no -2 suffix collision.

It:

• Binds UDS at /tmp/cc-socks/<publish_pid>.sock (0o600, dir 0o700)
• Writes matching ~/.claude/sessions/<publish_pid>.json + .key
• Accepts JSONL frames, acks with peer_message_status on a dial-back conn only (never on the inbound conn)
• Logs raw + parsed (auth redacted) to stdout and appends to ~/.agent-bus/captures/<pid>.jsonl
• Unlinks only its own socket, session file and key on SIGINT/SIGTERM

A listener is started for every non-Claude kind (grok, omp, codex, other) whenever session_start() runs, which in practice means whenever agent-bus mcp starts. Claude sessions are the one kind that never get one, since they already have their own socket.

agent-bus send <name> -m "text" reaches a Claude peer over UDS; the transport is chosen from the target's kind, so there is no vendor-specific send command to remember.

Claude users: install nothing. If a peer is running listen, it appears in /list-agents and messages arrive as cross-session.

CRITICAL SAFETY

• Received content must never auto-execute. Require explicit user approval.
• send-uds (not send) is the raw-frame experiment for reversing the wire format. Do not point it at anything except a socket we started with listen under test overrides.

Test overrides: AGENT_BUS_SOCK_DIR, AGENT_BUS_SESSIONS_DIR, AGENT_BUS_HOME.

Installation

Note: package name is agent-bus-team; the CLI is agent-bus.

# run the latest version with uvx
uvx --from agent-bus-team agent-bus

# or install with pip
pip install agent-bus-team
agent-bus

Running from source

gh repo clone danbarua/agent-bus && cd agent-bus
python -m pip install -e .
agent-bus --help

For Claude or omp: use python -m agent_bus or the installed agent-bus CLI. No plugin is required for Claude.

Installing it

pip install agent-bus-team        # or: uv pip install -e .

That is the whole of it. A peer joins the bus by running the MCP server — agent-bus mcp calls session_start() on startup and session_end() on exit, in-process, which registers the session and publishes its UDS listener. Point your harness's MCP config at it.

There is no plugin. This repo used to carry a plugin.json, hook scripts and a CLI wrapper for grok plugin install; all of it existed to run agent-bus from Grok's Bash tool, which is exactly what the MCP server removed. A plugin manifest that wires up nothing on install only advertises a capability that is not there. Whether a plugin is the right shape at all is still open — see hooks-in-foreign-harnesses.md for why the last attempt was worse than nothing.

agent-bus hook session-start|session-end remains as a CLI entry point for a harness that has hooks and no MCP. It exits 0 always, never blocks on stdin, and writes only to stderr.

Claude Code: install NOTHING. No plugin, no .claude-plugin/, no .mcp.json, no skills or slash commands. Claude sees a Grok/omp/codex peer through native ListAgents/SendMessage because that peer's listen published the session file and socket.

The agent-facing surface is the MCP tool set above — the tool descriptions carry the consent rule. Use the CLI or import agent_bus.store for everything else.

Development / test

python -m pytest tests/ -q --tb=line
AGENT_BUS_HOME=/tmp/ab-test python -m agent_bus list --json

Limitations / non-goals

• No impersonation of Claude's full protocol — listen and the claude transport implement the UDS peer messaging subset only (see UDS-protocol.md).
• No auto-start of other agents.
• Herdr TTY injection is a separate channel (not used here).
• Inboxes are per bus-home; multiple users would need separate homes or sync.

This is intentionally small and boring.