pi-msg

Drive the Pi coding agent entirely from an XMPP chat client — 1:1 or in a group chat (MUC).

pi-msg launches pi --mode rpc, then bridges Pi's JSONL event stream to XMPP (via mellium.im/xmpp): the assistant's replies are relayed to you as chat messages, and your chat messages drive the agent — plain prompts and slash commands, exactly as if you'd typed them into Pi locally.

Because it runs Pi in RPC mode, commands like /new work over chat (an earlier in-process-extension version couldn't do this — sendUserMessage can't invoke Pi's command layer).

How it works

Conversations are persisted across restarts: pi-msg records the pi session file (<config-dir>/<account>.session) on startup, whenever the session changes (/new, /resume, /fork), and on shutdown — then resumes it on the next launch via pi --session <file>. A bridge restart therefore continues the previous conversation; only /new resets context (or an explicit --prompt on-demand spawn — see below). If the saved session file is missing or empty, pi-msg starts a fresh session instead.

sequenceDiagram
    participant You as You (XMPP client)
    participant Bridge as pi-msg
    participant Pi as pi --mode rpc
    You->>Bridge: "fix the build"
    Bridge->>Pi: prompt
    Pi-->>Bridge: message_end event
    Bridge-->>You: assistant text
    You->>Bridge: "/new"
    Bridge->>Pi: {type:"new_session"}
    Note over Pi: fresh session

• Each finished assistant message → sent to you as chat.
• Agent state shows on three independent signals (1:1): a typing indicator while a reply is actually being written, presence <show> (dnd while busy, available when idle), and a presence status label of the current activity (thinking…, running: <cmd>, replying…, retrying…, listening). When a run settles with no text you get a ✅ done (no reply) — your turn nudge.
• Messages you send are acknowledged with read receipts — XEP-0184 delivery receipts and XEP-0333 chat markers (displayed) — when the agent takes them in, if your client requests them.
• Your chat messages → routed to Pi:

Every bridged command also works with a ! prefix — /new and !new are interchangeable. ("/", "!") only matters for the owner: non-owners' messages are always treated as literal text.

Configuration

Create ~/.config/pi-msg/config.json (override the path with PI_MSG_CONFIG), then chmod 600 it:

{
  "accounts": {
    "default": {
      "jid": "pi@chat.example.com",
      "password": "super-secret",
      "owner": "you@chat.example.com",
      "model": "anthropic/claude-sonnet-latest",
      "workdir": "/path/to/your/project"
    }
  }
}

Per-account fields:

Multiple accounts: add more keys under accounts; default is used unless you set PI_MSG_ACCOUNT=<name>. In 1:1 mode only the owner JID may drive the agent.

Group chat (MUC)

Set room on an account (a single MUC JID, or an array of them) and pi-msg also joins each. The owner's 1:1 stays the primary channel — joining a room is purely additive and doesn't change 1:1 behaviour (lifecycle notices, and unsolicited output all still go to the owner). The typing indicator now tails the reply's to: routing line (issue #44): it points at whichever 1:1 recipient the reply names — the DM, another agent — and stays dark when the reply heads to a room or to: noop. Each reply goes back to wherever its routing line points, including the specific room when several are joined. Room messages are handled on two independent axes:

• Trigger — does the message start/steer a turn?
  • the owner → always
  • anyone else who addresses the bot by name (pi: … / pi, …) → always
  • all other chatter → never (it's buffered as ambient context)
• Authority — is the content trusted?
  • the owner → canonical (authoritative)
  • everyone else, even when addressing the bot → untrusted commentary; the agent is told to use its judgment and is under no obligation to act on it

Untriggered messages are buffered and, on the next turn, prepended to the prompt as a clearly-labeled "room commentary — non-canonical" block, then the buffer clears.

Reply routing (explicit from:/to:). When an account has room access, routing is fully explicit — no guessing. Each prompt the agent receives leads with a header naming the message's origin:

from: <channel jid>     # the room (group msg) or the owner (DM) — reply here to answer in place
sender: <person jid>    # room messages only, when the real JID is known — reply here to DM them
<message body>

And every agent reply must begin with a to: <jid> line naming its destination:

• to: <room jid> → the group chat (groupchat)
• to: <owner or occupant jid> → that person, 1:1

One reply may contain several to: blocks — each to: line starts a new message, so the agent can fan a single turn out to multiple destinations:

to: team@muc.chat.zachmanson.com
Deploying now — back in 5.
to: zach@chat.zachmanson.com
(privately: the staging creds are stale, heads up)

Destinations are allowlisted: the owner, joined room(s), and real JIDs currently seen in a room. A reply whose to: is missing or points anywhere else is sent to the owner, so nothing is silently lost — the agent can't message arbitrary users. In a pure 1:1 account (no room) there are no prefixes; replies just go to the owner.

File transfer. The agent sends files with the send_file tool (a structured tool call, not in-band text — see Agent tools below): pi-msg uploads the file via XEP-0363 HTTP Upload and sends the resulting URL as an XEP-0066 out-of-band message, so the recipient's client shows a downloadable file. The destination is allowlisted (owner, joined rooms, known occupants) exactly like a to: reply. The upload component is discovered automatically (upload.<domain> / httpupload.<domain>) or set explicitly via the uploadService config field.

The room must be non-anonymous (ejabberd: "Present real Jabber IDs to → anyone", optionally members-only). The owner is recognized by real JID; in a semi-anonymous room real JIDs are hidden, so the owner can't be distinguished and every message falls through to the untrusted/ambient tiers.

Errors dumping ground (errorRoom). Set errorRoom to a bare MUC JID (e.g. errors@muc.chat.example.com) and pi-msg uses it as a write-only dumping ground for agent replies it can't route (no to: line, text before the first to:, or a non-allowlisted destination). This lets you mute the room and only check it when you need to recover something — without the dropped content spamming your 1:1.

The bridge joins the room at the XMPP layer (so groupchat sends are accepted and the keepalive covers it), but deliberately keeps it out of the agent-visible room set: it is never dispatched to the agent, never appears in the reply/file allowlist, and isn't tracked for occupants. So the agent can't read the room or route anything to it — it's write-only by construction, which keeps multiple agents from acting on each other's rejected output. If errorRoom is unset, unrouteable replies fall back to the owner's 1:1 as before.

Agent tools

Beyond reply text, the agent gets structured tools (registered by a small companion extension that pi-msg loads into pi --mode rpc, which relays each call back to pi-msg to perform the XMPP action):

Reply routing (to:) stays an in-band text convention (above); only these discrete side-effect actions are tools.

Run

go build -o pi-msg . && ./pi-msg     # from the repo

Nix

nix run   github:zachpmanson/pi-msg    # run the bridge
nix build github:zachpmanson/pi-msg    # build the package (bin: pi-msg)

Dev shell (Go + gopls) via nix develop, or automatically with direnv — the repo ships a .envrc (use flake); run direnv allow once.

Set PI_MSG_DEBUG=1 to print connection/status/stderr diagnostics. On startup the bot simply comes online in your roster (presence listening); on shutdown or a pi crash it goes offline with a <status> describing why and when — pi-msg no longer posts chat banners for these lifecycle events.

On-demand spawns: --prompt

pi-msg --prompt "<task>" (alias --command) spawns a fresh, on-demand persona with the task as its very first prompt — no separate XMPP-send hop needed to wake it. It intentionally does not resume the saved session (stateless by construction) and skips restart-gap replay; the reply routes to the owner per the normal routing contract. This backs the sentinel doer flow (zachpmanson/beltino#18).

The same payload can ride the existing one-shot start-directive file (<config-dir>/<account>.start, written via writePromptDirective):

prompt
resolve zachpmanson/pi-msg#35 and open a PR

Either way the directive file is consumed (one-shot); an explicit --prompt flag overrides a file-delivered payload. Routine restarts that carry no prompt keep the existing resume + proactive/idle behavior unchanged.

Requirements: Go ≥ 1.26 (to build), and a pi on PATH that's logged into a provider (pi → /login).

Notes

• Pi runs tools autonomously (no built-in approval prompts). If some other extension raises a dialog (select/confirm/input/editor), pi-msg auto-dismisses it (nobody's at the TUI) and tells you over chat — so approval-gated tools are declined over the bridge.
• Auth uses SASL SCRAM-SHA-256 (mellium negotiates it cleanly against ejabberd); STARTTLS is required first.