agentcairn

The command line for Agent Cairn: shared, persistent project memory for AI coding agents, stored in Neo4j.

Connect Claude Code (one command, every project)

You need a running Agent Cairn server and a token from its dashboard (Access → Create token). Then, once per machine:

claude mcp add cairn --scope user -e CAIRN_URL=http://localhost:8787 -e CAIRN_TOKEN=<token> -- npx -y agentcairn mcp

Check it with claude mcp list (it should say Connected), then restart Claude Code. If it says "MCP server cairn already exists in user config", run claude mcp remove cairn --scope user and add it again.

agentcairn mcp is a small bridge to your server. It takes the project from the folder Claude Code was started in (its .cairn.json, else the git remote), so there is nothing to configure per repository.

Commands

npx -y agentcairn <command>     # or: npm install -g agentcairn, then use `cairn`

cairn init [project-id]    # link this folder to a project; add hooks and a CLAUDE.md block
cairn doctor               # check the whole setup
cairn briefing             # print what an agent would be told
cairn resume               # everything needed to carry on where the last agent stopped
cairn remote <git-url>     # make another git URL resolve to this project
cairn install-command      # add the /cairn slash command to Claude Code
cairn mcp                  # the stdio bridge used above

cairn init writes .cairn.json, a .mcp.json server entry (only when the server is not already registered for every project; no secrets, the address and token come from your environment), Claude Code hooks and a short CLAUDE.md block. The hooks inject the briefing at session start (opening with where the last work stopped), record the session as it goes (what was asked, files, commits, uncommitted work; set CAIRN_SESSION_CAPTURE=meta or off to send less), remind the agent to write a handoff after substantial work, and stamp memories with the commit and author. Set CAIRN_URL and CAIRN_TOKEN in your environment for the hooks to reach the server.

Everything here runs on your machine, because the server cannot see your repository. See Using it with Claude Code. MIT licensed.