Skip to content
Docs · Any MCP agent

DocsBuild with it

Use ArcRoom memory from any MCP agent

ArcRoom's tools are a small MCP server that talks over stdio. Any agent that can start an MCP server can use them, even onearcroom init doesn't set up. You get the tools only; the automatic parts listed below don't run.

1. Run arcroom init once

Run it once on this laptop first. It saves your key and installs the memory server in ~/.arcroom/runtime. Init also writes Claude Code's settings in this folder (.claude/settings.local.json); if you don't use Claude Code here, that file does nothing.

2. Add the server to your agent

This is the same command arcroom init registers for Claude Code. Put your own home folder in the path; on Windows use C:/Users/you/.arcroom/runtime/index.js. Most agents take this shape in their MCP settings:

{
  "mcpServers": {
    "arcroom": {
      "command": "node",
      "args": ["/home/you/.arcroom/runtime/index.js", "mcp"]
    }
  }
}

It needs Node 20 or later. Start the agent in your project folder: the server reads the git branch and file hashes from the folder it starts in.

What you get

  • find_task and split_task to look up work a teammate's agent already did, open_task to open a part made of its own parts when the agent starts it; claim_task and complete_task to save new work.
  • remember for decisions, know for what the team knows about a topic, propose_rule to suggest a rule a person accepts on the Rules page, fail_task with the number of the step that broke, so the next agent redoes the work from that step, and the version tools (pin_task, unpin_task, rollback_task, evict_task).
  • A lookup waits at most 4 seconds; after that the agent is told nothing was found and carries on.
  • arcroom off and a revoked key apply here too: the tools answer with an error and send nothing.

What you don't get with the tools alone

  • No hand-over on each prompt: the agent has to call find_task itself.
  • No session reports, no note from the last session, no loop guard, no done-checks.
  • Code changes aren't recorded on their own. The server tells your agent to save finished work, code changes included, with complete_task.

For Claude Code, Codex, Cursor and OpenCode, use arcroom init --agent … instead, which sets up everything together. See Agent support.