Local MCP relay for cross-model agent teams. Your Claude, Codex, and other agents talk to each other over Unix sockets. No broker, no network, no accounts.
Let your AI agents talk to each other. Relay is a local messaging protocol for agent teams. Picture a coder in one pane and a verifier in another, running on different models under different harnesses, both working the same task. It runs on your machine. No server, no cloud.
The second line arms the agent with loop instructions, a wake hook, and tool trust, so conversations run unattended. There's more on that under Arming your agents. setup codex does the same for Codex and registers the server too. For any other MCP harness, the server command is npx -y @modulastack/relay --name . To embed it in your own tool, npm install @modulastack/relay.
1. Open two terminals in the same project folder. 2. Start an agent in each. Give one --name coder, the other --name verifier. 3. Tell either agent: "check the relay roster." You'll see the other peer right away. The roster pings peers directly, so it doubles as your "is the link up?" check and answers in milliseconds. 4. Now tell the coder, in plain English:
Send the verifier this message over relay: "please review utils.py and reply with any issues", and wait for the reply.
5. Watch the verifier receive it, do the work, and reply. You never type protocol commands. The agents drive Relay themselves.
Expect the speed to come from the models, not the link. Each exchange is a full turn on each side, since the other agent is actually working on your request, so replies land at whatever pace the model runs, usually a few seconds. If it drags, that's the model or a rate-limited account, not the transport. relayroster checks the link b
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/modulastack--modula-relay.html)
MCP-compatible context and memory system: paged code navigation for AI assistants, plus a recorded, verified theory of what is known about the code.
luckeyfaraday/frontier-orchestratorMulti-agent orchestration for Claude Code: MCP server + skill that keeps Claude as lead engineer while delegating backend work to OpenAI Codex and frontend/design work to Kimi
menot-you/n-memoryMemory your agent can trust — every answer carries its source; when it doesn't know, it says so. Hermetic, local, MCP.
thaildhe172591/pythiaPL/SQL development for AI agents on Oracle Database - expert queries, impact analysis, and a snapshot-verified write path with honest rollback
0xwilliamortiz/openclaude-improvedruns anywhere. uses anything
howznguyen/pi-delegate-mcpMCP server that turns the pi coding agent into a steerable background worker: delegate a task, redirect it mid-run, and keep its context out of your own.
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.