crossagent MCP server

crossagent

crossagent pairs two coding agents running in adjacent tmux panes: one drives, the other reviews. Every plan and diff goes through the peer before it ships, and the reply always ends in a verdict. tmux is the only transport — no network, no shared file, no socket.

Install

Requirements: Node.js 22 or newer, tmux, and two panes each running a coding agent (Claude Code, Codex CLI), both inside the same git repository for automatic peer discovery. Each agent must call ping once — registration happens on that call, not on server startup.

Run npx from outside a clone of this repo — inside it, npx resolves the local package instead of the published one.

Step 1. Add the MCP server to each agent.

claude mcp add crossagent -- npx -y crossagent-mcp
codex mcp add crossagent -- npx -y crossagent-mcp

Step 2. Link the pair skill so both agents load the pairing protocol proactively.

npx -y crossagent-mcp setup --link

This symlinks skills/pair into ~/.claude/skills/pair, ~/.codex/skills/pair (or $CODEX_HOME/skills/pair) and ~/.agents/skills/pair. setup --unlink reverses it, removing only the symlinks it created.

Without step 2, the protocol is still available as the pair MCP prompt (/mcp__crossagent__pair in Claude Code). Details in docs/skill.md.

brief sends the harness's reset command to the peer pane (/clear for Claude Code, /new for Codex), clearing its current conversation. It is never needed for discuss to work and never runs automatically.

How it works

• ping — each agent registers its pane, harness and repository so the peer can be found.
• brief — resets the peer's conversation. Call it by hand only to clear a stale peer; it sends no role text.
• discuss — every round's task and context go here. Each message carries a short role brief (same-level copilot or reviewer bound to the VERDICT: contract), so the peer needs no setup, and every plan or diff sent this way ends in one verdict line.

The last line of every review is VERDICT: <word>:

Verdict	Meaning	Driver does
AGREE	No corrections.	Ships it as is.
ADJUST	Fine once the listed items are fixed.	Fixes them, or argues back when one is wrong.
OBJECT	Do not proceed as written, reasons included.	Reworks the plan and asks again.
ESCALATE	Only the human can call this one.	Stops and brings it to the user.
UNPARSED	Not a verdict — reply had no readable VERDICT: line.	Asks again.

The full round format, review extraction and verdict parsing are in docs/protocol.md.

Tools

Tool	Description
ping	Registers this pane's identity and returns pane/harness/repo/mode.
list_peers	Lists other crossagent-visible tmux panes.
send	Sends literal text + Enter to a tmux pane.
read	Captures the last N lines from a tmux pane.
wait_reply	Waits until the peer pane's harness looks idle, returns the new content.
brief	Resets the peer pane's conversation (/clear, /new).
discuss	Sends a review round (role brief + message), parses the VERDICT: line.
ask_review	Alias of discuss with kind: "implementation".
pair (prompt)	The full pairing protocol as a single user message.

Arguments and defaults: docs/tools.md. When pane is omitted, the target is the repo-scoped agent peer: shell panes are ignored, and with several agents the one in your own tmux window wins if it is the only one there.

Configuration

Optional config lives at ~/.config/crossagent/config.toml. Each harness's idle/busy detection regexes and its reset command (used by brief, e.g. /clear for Claude Code, /new for Codex) are overridable per harness, with tunable defaults for timeout, quiet period, and tail-window size. See docs/configuration.md for the full reference and examples.

How it finds the other pane

Each pane resolves its own tmux pane id, harness, and repo identity when ping is called, then self-declares them as tmux pane options so list_peers can find it. Codex CLI starts MCP servers lazily, so a Codex pane may not be discoverable by its declared identity until its first tool call. See docs/discovery.md for the full resolution logic and fallbacks.

Contributing

Bug reports, harness support, and protocol improvements are welcome. See CONTRIBUTING.md for dev setup, tests, and how to add a new harness. This project follows the Contributor Covenant.

Using another coding agent? Adding a harness means a detection entry in src/harness.ts and a config block; see CONTRIBUTING.md.

License

MIT