agents-control

Stdio MCP that lets one coding agent operate another: status, redacted config, allowlisted-or-confirmed writes, logs, doctor/fix, debug, CLI, gateway restart.

Same Mac, same user. This is a structured control plane, not a new security boundary.

Peers

npx agents-control --list-peers

--peer openclaw is what Hermes should run (to control OpenClaw).
--peer hermes is what OpenClaw should run (to control Hermes).

Related MCP server: Engineering Knowledge Graph MCP Server

Tools

On peers without a config/doctor CLI, those tools return a short “use cli” message instead of failing the whole server.

Install

git clone https://github.com/devedbox/agents-control.git
cd agents-control
npm install

Hermes → OpenClaw (~/.hermes/config.yaml)

mcp_servers:
  openclaw-control:
    command: node
    args:
      - /ABS/PATH/agents-control/src/index.js
      - --peer
      - openclaw
    timeout: 180

OpenClaw → Hermes

openclaw mcp set hermes-control \
  '{"command":"node","args":["/ABS/PATH/agents-control/src/index.js","--peer","hermes"]}'

Add more servers the same way (--peer opencode, --peer claude, …).

Policy

• Reads are redacted (token / password / apiKey / …).

• Secret writes, doctor --fix, restart, uninstall, --no-redact require confirm=true.

• cli is execFile(peer, argv) — no bash -c.

License

MIT