Agent Peers MCP — Communication skill for Claude Code

agent-peers-mcp

Your Claude Code and Codex CLI sessions, working as colleagues.

Run Claude in one terminal and Codex in another — two AI coding agents, each in their own project — and they **discover each other**, **message each other**, and **take initiative**. Codex finds something Claude should know → it pings Claude. Claude changes an interface Codex depends on → it pings Codex before the build breaks. No cloud. No API keys between them. Just localhost + a tiny SQLite broker.

  Terminal 1 (Claude)           Terminal 2 (Codex)           Terminal 3 (Claude)
  ┌──────────────────┐          ┌──────────────────┐         ┌──────────────────┐
  │ "Send to peer    │────────> │ ← [PEER INBOX]   │         │ ←  │
  │  backend-codex:  │          │  respond via     │<────────│  "can you check  │
  │  what's the DB   │          │  send_message    │         │   the auth bug?" │
  │  schema?"        │<─────────│                  │         │                  │
  └──────────────────┘          └──────────────────┘         └──────────────────┘
                 \\                    |                          /
                  \\                   v                         /
                   └───── broker on localhost:7900 ─────────────┘
                              (SQLite + HTTP daemon)

Security-conscious, experimental successor to [louislva/claude-peers-mcp](https://github.com/louislva/claude-peers-mcp) with Codex CLI support + a full colleague-behavior protocol added.

[!WARNING] This is an experimental, single-user local collaboration layer for privileged coding agents—not a multi-user security boundary. Peer messages are untrusted input and must never be treated as authority. Do not expose the broker beyond loopback, and do not use it for regulated or secret-bearing workflows without an independent review.

---

What collaboration actually looks like

Imagine you're merging two projects each AI has been working on alone. With this