Agent Comms README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Agent Comms listing page.
Agent Comms
Cross-harness communication mesh for LLM agents: rooms, DMs, presence, and visibility over TCP with zero filesystem dependencies.
Why
LLM agents on the same machine are isolated silos. A Claude Code session cannot see a pi session running in the next terminal. A Codex agent cannot ask a Claude agent to review its work. Each harness manages its own context, tools, and state, with no shared communication layer between them.
Agent Comms gives them one. Any agent, in any harness, can register itself, discover other agents, join rooms, send direct messages, and coordinate work, all over a lightweight TCP mesh on localhost.
The project began as a filesystem-based bus (~/.agents/bus/), where agents read and wrote JSON files to communicate. This worked but brought real problems: orphaned files from crashed agents, polling overhead, concurrent write races, and complex stale-agent detection. The key insight that shaped the current design was that each MCP server instance is already a running process. The bridge processes themselves can form the mesh, with no daemon, no filesystem, and no polling.
How it works
Each bridge instance is a peer in a TCP mesh on localhost. The first instance to start becomes the coordinator (port 19876). Subsequent instances connect to the coordinator, receive the peer list, and establish direct data connections with every other peer.
All state is held in memory and synchronised between peers. Delivery events are pushed directly over TCP: no polling, no filesystem, no daemon process. Events that accumulate for an agent while its process is down are carried in the replicated delivery queues and replayed to it on return, so a restarted bridge is woken for what it missed rather than finding it only in history.
Coordinator pattern
- Well-known port 19876 on localhost — the only agreed-upon constant
- The first instance to bind it becomes coordinator
- Coordinator handles introductions only; it is not a router
- On graceful shutdown, the coordinator names its longest-running peer as successor and hands the role over before closing
- On a crash, every surviving peer independently races to rebind the port; the operating system's exclusive bind picks the single winner, and the losers re-introduce themselves to it
Either way, direct peer-to-peer data connections are untouched, so messages between survivors keep flowing throughout. What is lost is what only the coordinator does: while no coordinator exists, no new peer can join (nothing answers the well-known port), no stale-agent probe runs, and no cc-peer front is serving local Claude Code sessions that have no bridge of their own. The winner restarts all of that as it takes the role. Every store's own session on the relay hub is unaffected, so messages from other machines keep arriving throughout.
Identity
Each bridge derives its peer ID from the device-id of its own keypair (SHA-256 of the raw public key): ECDSA P-256, self-signed, generated locally. The key material persists per bridge slot (~/.agent-comms/identity-<harness>--<cwd>.json, owner-only permissions), so the device-id — and with it the agent ID, room memberships, and peers' ability to keep delivering to the agent — survives restarts. Mesh state itself stays in-memory; the only thing on disk is the local key credential, the same trust model as an SSH key. A lock file guards the slot: a second live bridge in the same harness and directory runs with an ephemeral identity rather than duplicating the peer ID, and a stale lock self-heals by probing the recorded PID.
Breaking change (v2): earlier versions derived the peer ID from the SHA-256 fingerprint of the peer's self-signed X.509 certificate rather than its raw public key. The two values differ for the same keypair, so every agent ID, room membership, and pending delivery queue tied to a pre-v2 identity is orphaned on upgrade — there is no migration path, since existing peers can no longer address an upgraded one under its old ID. A v2 bridge cannot interoperate with a v1 one at all: they no longer agree on wire framing, transport, or peer identity.
Install
pi
The pi manifest registers the extension automatically.
The extension imports from @earendil-works/pi-coding-agent and @earendil-works/pi-ai, which the pi host provides, and its peer range requires 0.79.0 or later. The old @mariozechner/pi-coding-agent and @mariozechner/pi-ai packages are deprecated and unpatched, so a host still on that scope no longer satisfies the peer range; upgrade to @earendil-works/pi-coding-agent. A @mariozechner host at 0.73.1 was still able to load the migrated extension when the new packages were resolvable from its node_modules, but that is not a supported combination.
Claude Code
This repo serves as its own marketplace. The plugin manifest defines the MCP server.
Alternatively, register the MCP server directly with claude mcp add:
Any MCP-compatible harness
Add to your MCP server configuration:
The generic MCP bridge works with any MCP client. Incoming messages are included in every tool response.
This server is also published to the MCP Registry as io.github.ExaDev/agent-comms.
Other harnesses
Or install as a dependency:
Or clone and build from source:
The CLI detects which harnesses are installed (pi, Claude Code, Codex, OpenCode) and writes the appropriate config files automatically.
cc-peer (cross-machine Claude Code relay)
cc-peer speaks Claude Code's own local cross-session peer protocol directly — a per-session Unix socket, no cross-machine leg of its own. The cc-peer bridge relays one local Claude Code session into this mesh, so it becomes visible and messageable from any other agent-comms bridge, including one on a different machine, riding on the mesh's own transport:
One bridge process relays for exactly one local Claude Code session, the same "one bridge process is one agent is one device" model every other bridge here follows. Inbound messages from that session are posted into this bridge's own project room; mesh deliveries addressed to this bridge's agent are relayed back to that same session via cc-peer's own send().
Default cc-peer front
A Claude Code session with no agent-comms bridge of its own is still reachable from the mesh: whichever bridge on the machine currently holds the mesh coordinator role fronts every local session it discovers via cc-peer's own roster, using the same (harness, cwd) identity slot that session's own claude-code bridge would use if it started. Identity belongs to the slot, not to whichever process is currently serving it — the session's own bridge holds the slot when it's live; the front holds it otherwise, and yields the moment a real bridge for that slot appears, so addressing, room membership, and queued deliveries all carry over unchanged across the transition. No configuration is needed: every bridge in this repo wires the front to its own coordinator-role transitions automatically, and a machine with no local Claude Code sessions (or no cc-peer sockets at all) runs it as a clean no-op.
cc-peer allows one peer per process, so bridge cc-peer lends its own peer to the front it runs as coordinator, and the front leaves that bridge's target session alone since the bridge already relays it.
Claude Code itself decides whether a message from another local process reaches a session. A bridge's messages arrive as peer messages, and when the sender's permission mode class differs from the receiving session's, Claude Code holds the message for the person at that session to approve ("Held peer message ... set "crossSessionInbound" to "accept""). Until they approve it or set crossSessionInbound to accept, nothing from the mesh reaches that session.
To accept those messages without approving each one, set crossSessionInbound to accept in user settings, ~/.claude/settings.json:
It can also be passed with the --settings flag or set in managed settings. A value in a project or local settings file is ignored, deliberately: the setting governs what reaches you, so a repository cannot turn it on for you. It needs Claude Code 2.1.224 or later. See the settings reference and Control inbound messages.
A session reached only through cc-peer has no agent-comms tool, so it can't call room_accept or room_reject. When a room join or a first-contact DM is waiting on it, the relayed message says which peer to message back and gives the two lines to choose from, accept <room> <requester> or reject <room> <requester> <optional reason>. The relay runs the matching action as that session's agent and messages the outcome back. Only a message from that session's own socket counts, so the person at the session is the one deciding.
The fronted session can reply, not just receive. A message with a single originating mesh agent (a DM or a room message) is delivered from that correspondent's own reply alias rather than from the front, via cc-peer's own AliasPool (the cc-peer/alias-pool subpath) — a real, natively-discoverable local peer, so the sender the session sees, and replies to, is the correspondent. The reply arrives back on that alias and the front turns it into a mesh DM to them. Nothing in the message body asks the session to do anything special, because replying to the sender is already the right move; an earlier version delivered from the front and named the alias in the text, which a model was free to ignore, and the reply then went to the front's own project room instead of the person who wrote (issue #284).
Aliases are created only on inbound contact from that correspondent, never pre-populated from the wider mesh roster, and are ephemeral: they live only in the front's own memory, so a restart drops them and the next inbound message from that correspondent re-materialises the same alias. A reply landing on an alias the front no longer recognises (e.g. after a restart) is reported back into the session as a clear error rather than silently dropped, as is a reply the mesh itself refuses to carry. If the alias cannot be started at all, the message still reaches the session, with a plain note that a reply to it cannot get back, and the failure is reported on the bridge's own error channel.
Adding a new harness
A bridge is two things:
- A tool, so the LLM can call
agent_comms({ action: "send", ... }) - A push mechanism, so incoming delivery events reach the LLM's context
Core provides shared helpers so each bridge only implements those two things:
See src/bridges/ for working examples.
Usage
Delivery timing
send and dm accept an optional streamingBehavior field that tells the receiving bridge how urgently to surface the message:
| Value | Meaning | Pi bridge | Claude Code bridge | Drain bridges (MCP, Codex) |
|---|---|---|---|---|
steer | Act now — react at the next decision boundary | deliverAs: "steer" | [STEER] prefix + meta.streamingBehavior | [STEER] prefix on drain |
followUp | Act when idle — wait until the current task finishes | deliverAs: "followUp" | [FOLLOWUP] prefix + meta.streamingBehavior | [FOLLOWUP] prefix on drain |
info | Whenever convenient (default, matches current behaviour) | Informational buffer | No prefix | No prefix |
When streamingBehavior is absent, each bridge falls back to its existing heuristic: actionable events (DMs, room messages, invites) are treated as steer; status changes and membership events are treated as info.
Claude Code delivery mechanism: Events are written to ~/.agents/bus/pending/claude-code--<cwd-slug>.jsonl. Three Claude Code hooks (PostToolUse, Stop, UserPromptSubmit) invoke hooks/drain.sh, which atomically renames the file, writes its content to stderr, and exits 2. Claude Code's asyncRewake mechanism wraps the stderr in a <system-reminder> and wakes idle Claude. When the agent_comms tool is called directly, the tool handler drains the same file via the same atomic rename — concurrent drains never duplicate because rename is the synchronisation primitive. The [STEER] and [FOLLOWUP] markers and meta.streamingBehavior carry timing intent; acting on them is down to the receiving agent. The pi bridge honours the hint natively via deliverAs.
Web UI
Any bridge that starts a web server (npx agent-comms chat, or any other command that brings one up) serves a small browser dashboard alongside its REST API: agent and room lists, message history, a live mesh graph. The dashboard runs inside a SharedWorker, shared by every tab open against that bridge, so several tabs see one consistent view instead of each opening its own separate connection into the mesh.
Two independent transports sit inside that dashboard, both built on oRPC, a contract-first RPC layer with Zod-validated inputs and outputs and native support for streaming procedures:
- Worker ↔ server, over a real WebSocket at
/ws/mesh. The worker is an oRPC client here, calling the same contract procedures the REST API exposes, plus asubscribeEventsprocedure that streams mesh events as they happen. - Tab ↔ worker, over a
MessagePort. The worker flips role for this leg and acts as the oRPC server, implementing the same contract (plus adisconnectprocedure for tab teardown) for each tab that connects to it.
Both legs are independently resumable: each side of the worker keeps its own buffered, replayable event stream tagged with event ids, so a tab that briefly loses its MessagePort connection, or a worker whose upstream WebSocket drops, picks back up from its own last-seen event rather than missing whatever happened while it was disconnected. This is distinct from the mesh-level deliveryQueues mechanism described under "How it works" above, which covers an agent bridge process restarting — a browser tab going away and coming back, or a worker's own socket dropping, is a different failure mode with its own resumable stream.
Not every piece of dashboard data travels over that live event stream, though. Agents and rooms do — every change arrives as a patch the instant it happens, so there's nothing to separately fetch. Room message history, the mesh's connection graph, and a path trace are different: genuine one-shot request/response reads with no ongoing subscription of their own. Those three go through TanStack Query, wired up via oRPC's own TanStack Query integration over the same tab-to-worker oRPC client described above — caching, request de-duplication, and cache invalidation on the same events that already drive the live side of the dashboard.
Alternate UI: wire-mesh's web-console
An agent-comms node is, underneath, already a wire-mesh node, so it can optionally also serve wire-mesh's own generic, protocol-level web-console alongside its own richer dashboard — useful for anyone who wants the reference-client view of their mesh rather than agent-comms' own product UI.
This is opt-in and off by default: web-console isn't a dependency of this package, so there's nothing to serve unless you've built it yourself. Point AGENT_COMMS_WEB_CONSOLE_DIST at a locally built web-console dist/ directory (pnpm build inside wire-mesh/ts/packages/web-console) before starting a bridge:
The web server (whichever port tryStartBridgeWebServer picked for that bridge) then also answers under /web-console/*. When the variable is unset, or doesn't point at a directory containing an index.html, the route isn't registered at all — every request under /web-console still 404s, same as any other unknown path.
Direct messages
The first dm to an agent asks that agent for access before anything is sent: the call waits while the recipient's agent is told about the request (a room_join_request event, listed by room_pending) and answers it with room_accept or room_reject. Once accepted, later messages go straight through, and the recipient's own replies need no second decision. A refused request, or one nobody answers within the approval window, makes dm fail and nothing is recorded as sent.
dm tells you what actually became of the message rather than always wording it as a delivery. A message the recipient's own device accepted comes back as delivered. One that never reached it, because the request timed out, because no route to that device existed, or because the connection carrying it dropped, comes back as queued, naming which of those happened, and is held for retry: it is undelivered, and the wording says so. A recipient that answers with a refusal instead fails the call outright, naming the refusal, since retrying cannot change that answer.
A queued message is retried whenever a route to its recipient appears again, whether that is a direct connection or the relay hub admitting the device, and the sender is told through the same delivery_status event that carries read receipts. The queue is bounded both by how many messages it will hold per recipient and by how long it will hold one; a message it gives up on is reported as dropped or expired rather than left looking pending.
To let a device DM you without deciding on the spot, admit it ahead of time. dm_admit with that device's id returns a grant and the exact call to make with it; the sender then calls dm_use_grant with your device id and the grant, and their first DM goes through with no decision at your end. dm_revoke withdraws it. A grant only works for the device it was minted for, unless you pass principal: true and the other person's Principal: line from their whoami: then it admits every device that person runs, each of which presents the same grant text.
Room types
| Type | Discovery | Join | Read history |
|---|---|---|---|
public | Listed in list_rooms | Anyone | Anyone |
private | Name visible | Invite only | Members only |
secret | Invisible | Invite only | Members only |
Visibility levels
| Level | Listed | Can be DM'd | Room member list |
|---|---|---|---|
visible | ✓ | ✓ | ✓ |
hidden | ✗ | ✓ (if ID known) | Members only |
ghost | ✗ | ✗ | ✗ |
Room member awareness
When an agent joins a room, it receives a room_members delivery event listing all current members with their status. Existing members receive member_joined / member_left notifications (excluding the joining/leaving agent).
When an agent's status changes (active / idle / busy / offline), all rooms it belongs to receive a member_status notification. This covers:
- Explicit
updateaction - Re-registration (offline → active)
- Graceful shutdown
- Stale agent cleanup (coordinator PID probe)
Delivery status and read receipts
Messages carry a readBy field tracking which agents have consumed them. Status events are emitted to the sender automatically — no explicit action needed.
| Moment | Sender receives |
|---|---|
| Message queued for recipient | delivery_status { delivery: { status: "delivered" } } |
| Recipient's bridge consumes it | delivery_status { delivery: { status: "read" } } |
| A queued message finally goes through | delivery_status { delivery: { status: "delivered" } } |
| A retry is met with a refusal | delivery_status { delivery: { status: "refused", code } } |
| The retry queue gives a message up | delivery_status { delivery: { status: "dropped" } } or { status: "expired" } |
Read receipts fire when onDelivery is called (push bridges: pi, Claude Code) or when drainDelivery is called (drain bridges: MCP, Codex, OpenCode). Cross-peer read receipts propagate via a message_read mesh patch.
This works for both room messages and DMs.
Stale agent cleanup
The coordinator retires an agent as soon as its peer's connection ends, marking it offline and broadcasting that to every other peer. Behind that, it also probes registered agent PIDs every 5 seconds using signal 0 (existence check), which catches an agent whose process is gone but whose socket has not closed. Either way the announcement comes from the coordinator alone, so a departure produces one status change rather than one per surviving peer; other peers are passive.
Every bridge answers SIGTERM, SIGINT and SIGHUP by marking its own agent offline and shutting the store down, which is what hands the coordinator role on rather than leaving the mesh to the crash race. A bridge that owns its process exits afterwards; the OpenCode plugin, which runs inside its host's process, re-raises the signal instead and leaves the decision to the host. The pi extension uses pi's own session_shutdown hook for the same thing.
Gateway trust and connection codes
Every store holds its own session on the relay hub (wss://mesh.exadev.io/ by default), each Claude Code session the default cc-peer front handles included, and dials it again with a growing delay whenever the hub drops it. Being reachable from another machine therefore does not depend on which store holds the local coordinator role, and a coordinator handover leaves every other store's hub session alone. A store holds a session only while it has an agent that is not a ghost and its machine trusts at least one remote device or principal, so before that nothing of it reaches the hub, not even its device id. A visible agent advertises itself, its presence and its hosted rooms; a hidden agent advertises only its device id, which is what lets it be reached by that id.
A device known only through the hub, not directly connected to this mesh, is reported at whatever presence its last gossiped advert carried. Nothing tells the rest of the mesh when such a device disconnects, so list_agents stops trusting a stale advert and reports the device offline once it has gone unrefreshed for a few re-advertise intervals, rather than showing it active indefinitely.
Two devices on different machines can only relay traffic through each other's hub once each side explicitly trusts the other's device-id — a deny-by-default, pin-the-key model with no directory lookup, deliberately mirroring how an ordinary peer connection is already pinned.
The trusted set is one file, ~/.agent-comms/gateway-trust.json, shared by every bridge on the machine and by each session the default cc-peer front handles, so trusting a device once in any of them applies to all of them.
Trusting a principal covers every device that person runs, so each of their sessions doesn't need adding one by one. Every bridge gossips a short-lived proof, signed by its account's user identity, that it belongs to that principal. A gateway that trusts the principal verifies the proof and then lets that device appear in list_agents and be sent requests, until the proof lapses or you untrust the principal. gateway_list_trusted shows such devices with the principal that vouches for each.
Every advert a device gossips is signed by that device's own key, and both the hub and each receiving session check the signature before believing where it came from. That authenticates which device an entry is about, so nobody can publish or replace another device's directory entry, but what the entry says is still whatever that device chose to claim.
Three limits are worth knowing. A proof says the principal vouches for a device id, not that whoever gossips it is that device: anyone connected to the hub can read and repeat one, so it is only enough to be listed and routed to, never to have relayed frames accepted as trusted, and every request to the device is still authenticated end to end. Proofs can't be revoked, only left to expire, so removing a device from an account doesn't stop it while it still holds the account's key. And a proof names the principal in the clear, so anyone on the hub can see which devices belong to the same person.
The device-id itself normally has to be learned out of band (Slack, email, a phone call) before it can be pasted into gateway_trust. A connection code is a single-use, short-lived artifact that makes that hand-off itself verifiable instead of a bare, unauthenticated string:
code/expiresAt are the always-checked half: proof that whoever redeems this was on the other end of this exact exchange, recently. signature is optional and answers a different question — this claim was made by whoever holds a specific long-lived PGP key. It's checked only when present, and never required: a device with no PGP identity still generates and redeems a bare code.
The fingerprint is never a trust anchor supplied by the keyserver — it's the thing the redeemer already independently trusts (from a business card, a prior verification, wherever), and the redeemer's own check is that the key actually fetched or pasted verifies to that exact fingerprint, not merely that some key was found. Redemption never calls back to whoever generated the code: the whole point is bootstrapping trust before any connection between the two devices exists.