relay-mcp

External working memory for multi-agent systems. An MCP server that keeps every AI agent — on any machine — in sync through a shared task board.

Why

When multiple Claude Code agents work across sessions or machines, context resets every time. relay_summary restores full system state in ~300 tokens, so no agent ever asks "what were we working on?"

Install

claude mcp add relay \
  -e RELAY_API_URL=https://agent-task-tracker.onrender.com \
  -e RELAY_API_KEY=your-api-key \
  -e RELAY_AGENT=claude-code \
  -- npx relay-mcp

Replace RELAY_API_URL and RELAY_API_KEY with your Relay backend credentials. Set RELAY_AGENT to a unique name per machine (e.g. windows-agent, macbook, claude-code).

Usage

Start every session with:

Call relay_summary to see current tasks.

The agent gets back something like:

total:12  in_progress:2  pending:5  blocked:1  done:4

ACTION NEEDED:
  [urgent] Deploy to prod — Push to Render after env vars confirmed

That's it. Full context in one call.

Tools

| Tool | When to call it | |------|----------------| | relay_summary | Start of every session — orients the agent without burning context | | relay_create | When starting a significant task | | relay_update | Progress updates, status changes | | relay_done | Task complete | | relay_block | Stuck — surfaces to human immediately | | relay_list | Full task list when you need details |

Environment variables

| Variable | Required | Default | Description | |----------|----------|---------|-------------| | RELAY_API_URL | Yes | — | Base URL of your Relay backend | | RELAY_API_KEY | Yes | — | x-api-key credential | | RELAY_AGENT | No | claude-code | Name to tag tasks with |

Self-hosting

The backend is Agent Task Tracker — deploy it to Render in one click, point RELAY_API_URL at your instance.

License

MIT