librocat

Persistent AI memory, over MCP. This package is the librocat MCP server: zero runtime dependencies, no model call. With no token it opens no network connection either — it reads your Open Knowledge Format (OKF) Markdown files into an in-memory search index and gives your agent sixteen tools to search, read, link-walk, write, rename, weed, keep a tag thesaurus, write finding aids, and read the Git history of your knowledge. Your agent is the cataloger — this server just holds the library. Log in (below) and the same binary becomes the client for librocat Cloud instead.

Recommended: the whole plugin (MCP + Agent Skills)

If your agent supports Agent Plugins (Claude Code, Cursor, Codex, VS Code, and others), one command installs the server above and the four Agent Skills that teach an agent to catalog well:

npx plugins add librocat/librocat

Skills + MCP, separately

Plenty of agents don't support Agent Plugins yet but do support Agent Skills and MCP — both far more widely adopted. Install both; the skills instruct the agent to call this server's tools, so skills alone do nothing and this server alone works without the cataloging guidance:

npx skills add librocat/librocat

{ "mcpServers": { "librocat": { "command": "npx", "args": ["-y", "librocat"], "env": { "LIBROCAT_BUNDLE": "./okf" } } } }

LIBROCAT_BUNDLE points at your OKF directory (default ./okf) — commit it to your own Git repo; the repo is the source of truth, not this server.

Want the same memory on every agent and machine, with a librarian?

librocat Cloud: the same tools, hosted. Libro, the library cat, catalogs what your agent leaves blank and files duplicates, retired concepts, and stray tags on a weekly review shelf for you to approve, with a revision on every write.

Upgrading does not change your MCP config. This same binary is the Cloud client too — add a workspace token (from the dashboard's Connect page) as LIBROCAT_TOKEN, or run librocat login and paste it in, and it switches to Cloud on its next start: the same 16 tools, served from your workspace instead of the local bundle (ingest_repo still reads this machine's disk, now writing into Cloud; reindex becomes a no-op, since Cloud has no local index to rebuild). librocat logout switches back. librocat login also offers to push an existing local library into a freshly-empty workspace in the same step, or run librocat push [dir] yourself any time — nothing is lost moving from Local to Cloud. Restart your agent after logging in or out; most agents cache the tool list at startup.

• Repository: https://github.com/librocat/librocat
• Site: https://librocat.dev
• Security: with no token, this server opens no network connection and sends no telemetry. With a token, it opens exactly one: to your Cloud workspace, nowhere else. The agent that calls it is the AI. See SECURITY.md.

Apache-2.0. © 2026 Laughing Hermit, Inc.