gknowledge-mcp

An MCP server for gknowledge wikis. Claude Code, Codex, Cursor and any other Model Context Protocol client can search, read and grow a linked Markdown wiki that an LLM keeps up to date, following Karpathy's LLM Wiki pattern.

Sources are compiled into the wiki once. Agents then read the wiki instead of re-reading raw documents on every question.

• Plain files. Markdown with [[wikilinks]] and YAML frontmatter, an index.md catalogue and an append-only log.md. The folder opens in Obsidian or any editor.
• Cited and incremental. Each new source updates only the pages it touches. Every page records its sources, and answers cite pages.
• Offline by default. A built-in heuristic engine ingests and answers with no model and no API key. For better synthesis, set provider, base_url and model in the project's settings/settings.json (or from the desktop app) to use Ollama or any OpenAI-compatible endpoint.
• No dependencies. Standard library only; pip install "gknowledge-mcp[pdf]" adds PDF import.

Install and register

claude mcp add gknowledge -- uvx gknowledge-mcp --init ~/Knowledge/my-project    # Claude Code
codex mcp add gknowledge -- uvx gknowledge-mcp --init ~/Knowledge/my-project     # Codex CLI

For Cursor or other clients:

{ "mcpServers": { "gknowledge": { "command": "uvx", "args": ["gknowledge-mcp", "--init", "/home/you/Knowledge/my-project"] } } }

--init creates the project folder the first time. Without it, the server only opens an existing project. You can also pipx install gknowledge-mcp and use gknowledge-mcp as the command.

Tools

The desktop app

This package is the headless part of gknowledge. The full app is a native GNOME application (GTK 4 and Libadwaita) with sources, wiki, assistant and a review queue in one window, and it opens the same project folders. Install it from the signed APT repository or Flatpak.

MIT licensed. Issues and contributions: github.com/openapi/gknowledge.