pi-claude-link

Two-way messaging between pi coding-agent sessions and Claude Code sessions.

A pi session running this extension shows up in Claude Code's /list-agents, and the two can message each other in real time — no daemon, no broker, no extra services. It works by speaking Claude Code's own cross-session messaging protocol (the mechanism behind /list-agents + SendMessage), so pi and Claude interoperate natively.

Inspired by pi-intercom (pi↔pi); pi-claude-link does pi↔Claude.

What you get

• Pi appears in Claude. Every pi session auto-registers as a peer — it shows in Claude Code's /list-agents, and Claude can SendMessage to it.
• Real-time inbound. A message from Claude is injected into the live pi session immediately (idle → starts a turn; busy → steers the current turn), and pi's reply is relayed back to the sender automatically.
• A claude-link tool for the pi model:
  • claude-link({ action: "list" }) — list reachable Claude sessions
  • claude-link({ action: "send", to, message }) — send; the reply comes back into this session
  • claude-link({ action: "ask", to, message }) — send and block until the reply, returned as the tool result
• /claude-link command to list sessions from the pi UI, plus a bundled skill so natural language ("message the other session…") just works.

Requirements

• pi coding-agent — npm i -g @earendil-works/pi-coding-agent (or have pi on PATH).
• Node ≥ 20.19 (22+ recommended). ⚠️ On older Node, pi itself crashes at startup with webidl.util.markAsUncloneable is not a function (a bundled-undici incompatibility). If you see that, run pi under a newer Node (nvm use 22). This is a pi requirement, not specific to this extension.
• Claude Code with cross-session messaging enabled. It's on by default in recent builds; if your Claude sessions don't appear in each other's /list-agents, start them with CLAUDE_CODE_HARBOR_KITE=1. (pi-claude-link auto-discovers Claude's socket directory — usually /tmp/cc-socks — and co-locates with it.)

Install

pi install git:github.com/alonw0/pi-claude-link
# once published to npm:
#   pi install npm:pi-claude-link
# for local development (from a clone):
#   pi -e /path/to/pi-claude-link/index.ts

Then start pi normally — the extension activates on session start. Verify with pi list (should show pi-claude-link) or /reload inside a running pi session.

Remove with pi remove pi-claude-link.

Usage

From pi → Claude (in a pi session):

list the claude sessions            → calls claude-link({action:"list"})
message claude-code-7b: build passes → calls claude-link({action:"send", ...})

or /claude-link to list. Replies arrive back in your pi session automatically.

From Claude → pi (in a Claude Code session):

/list-agents            → shows  pi-<dir>
SendMessage to pi-<dir>: "what's the test status?"

The message appears in the pi session in real time; pi's answer is relayed back to your Claude session.

Recommended safety setting

Cross-agent messages are untrusted peer input. To require explicit approval for each inbound message on the Claude side, set in ~/.claude/settings.json:

{ "crossSessionInbound": "hold" }

(accept delivers silently, refuse opts out.) See Security below.

How it works

A single in-process TypeScript extension (index.ts) plus a dependency-free port of Claude's wire protocol (claude-protocol.ts). No build step — pi runs TypeScript directly.

• session_start → bind a Unix socket at ‹Claude's socket dir›/cc-socks/<pid>.sock and write ~/.claude/sessions/<pid>.json, registering the pi session as a Claude peer.
• inbound (a type:"user" frame) → strip the <cross-session-message> envelope → pi.sendUserMessage(...) (real-time) + send a delivery receipt + record the sender. The sender's display name is resolved from Claude's registry so it matches /list-agents.
• agent_end → relay pi's reply back to the recorded sender(s).
• claude-link tool → list reads Claude's registry (live-filtered); send/ask connect to the target's socket and write a peer frame; replies route back to our socket and are injected.
• session_shutdown → unlink the socket and remove the registry entry.

There's no broker or daemon — Claude's session registry is the hub. Anything else registered in that hub is also visible to list.

Security

Messages between agents are peer input, not user authority:

• On the Claude side they arrive as origin.kind:"peer" and are subject to the crossSessionInbound gate (use "hold" to approve each one).
• On the pi side, injected messages are framed "from another agent, not your user" — the model is instructed to treat them as peer requests and not as your approval.
• Sockets are 0600 inside a 0700 directory: the boundary is your user account (a same-user process could already reach these).
• Do not wire this extension to external/automated inputs. It is a path for untrusted content to reach a permissioned agent — keep the input side to things a human sends.

Development / testing

Extensions are plain TypeScript run in-process (no build). The test/ harnesses drive a real pi rpc session end-to-end; run them under a pi-compatible Node:

# override how pi is launched if `pi` on PATH isn't on a new enough Node:
#   export PI_CMD="/path/to/node22 /path/to/pi/dist/cli.js"
node --experimental-strip-types test/reg-test.mjs     # registration + cleanup
node --experimental-strip-types test/roundtrip.mjs    # inbound relay + outbound tool

dev-run.sh launches pi with the extension loaded for interactive testing.

Enable debug logging with PI_CLAUDE_LINK_DEBUG=1 (or touch /tmp/pi-claude-link-debug.on); logs go to /tmp/pi-claude-link-debug.log.

Compatibility

Verified against pi-coding-agent 0.80.6 and Claude Code 2.1.224. The Claude side relies on its cross-session messaging protocol; if a future Claude release changes it, claude-protocol.ts is the single place to update.

More

• SECURITY.md — trust model and how to report a vulnerability
• CONTRIBUTING.md — layout and how to run the tests
• CHANGELOG.md — release notes

License

MIT