Agent Room Cli

Agent Room CLI

Local CLI room for a human, Claude Code, and Codex.

Chinese documentation: [README.zh-CN.md](README.zh-CN.md).

The design goal is simple: keep the original AI CLIs, put them in one shared terminal room, and make agents answer only when they are mentioned.

you> @cc review the retry design
cc> I see two risks...

you> @codex do you agree with cc?
codex> Mostly, but I would change the backoff cap...

you> @cc fold codex's point into the final recommendation
cc> Updated recommendation...

What This Is

`agent-room` starts a local HTTP/SSE room server. Human messages, Claude Code, and Codex messages share one transcript.

`agent-room run claude` and `agent-room run codex` launch the original CLI inside tmux. Each agent gets:

• MCP tools for send_message, catch_up, and who
• a room event relay
• mention-based trigger policy
• buffered context for messages it did not need to answer immediately

The package also exposes an `agent-room-mcp` bin. This is an **internal** MCP server launched automatically by `agent-room run`; it is not meant to be invoked directly.

This is closer to an OpenClaw-style local agent room than a Telegram bridge.

Security Boundary

• The room binds to 127.0.0.1 by default and has no application-level authentication. host and trio refuse a non-loopback --host unless you also pass --unsafe-no-auth.
• --unsafe-no-auth is an explicit escape hatch, not a secure remote mode: every peer that can reach the port can read the transcript, follow the SSE stream, register participants, and send messages. Prefer an authenticated tunnel that keeps the room listener on loopback.
• The API does not enable cross-origin browser reads, and JSON request bodies are capped at 64 KiB. These are containment measures, not authentication.
• JSONL transcripts and markdown archives remain sensitive plaintext under ~/.agent-room/. Pattern-based redaction is best effort and cannot guarantee that every credential o