LLM Wiki MCP

MCP server for LLM-maintained markdown wikis — a persistent, compounding knowledge base instead of per-query RAG.

Inspired by Andrej Karpathy's LLM Wiki pattern. The LLM incrementally builds and maintains interlinked markdown pages; you curate sources and ask questions; the wiki stays current because maintenance cost is near zero.

Three layers

Layer	Path	Who writes
Raw sources	raw/	You (immutable)
Wiki	wiki/{type}/{slug}.md	LLM via MCP
Schema	AGENTS.md	You + LLM

Plus index.md (catalog) and log.md (timeline), maintained automatically on writes.

MCP tools

Tool	Description
wiki_get_index	Read master index.md
wiki_list_pages	List pages, optional type filter
wiki_get_page	Fetch one page (frontmatter + body)
wiki_search	Full-text search with snippets
wiki_list_raw	List immutable source files in raw/
wiki_read_raw	Read a raw source file (text or binary metadata)
wiki_put_page	Create/update page + optional log/index/git commit

Quick start

1. Create your wiki from the starter template

cp -r examples/starter-wiki ~/my-wiki
cd ~/my-wiki
git init
git add . && git commit -m "Initial wiki scaffold"

2. Configure Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "llm-wiki": {
      "command": "npx",
      "args": ["-y", "@jlnkrth/llm-wiki-mcp"],
      "env": { "WIKI_DIR": "/path/to/my-wiki" }
    }
  }
}

Restart Cursor. Copy examples/starter-wiki/AGENTS.md conventions into your wiki (included if you copied the starter).

3. Local development (from source)

git clone https://github.com/jlnkrth/llm-wiki-mcp.git
cd llm-wiki-mcp
npm install
WIKI_DIR=/path/to/my-wiki npm run start:stdio

See TESTING.md for the full test checklist.

Remote deployment (HTTP + Docker)

For team/shared wikis on a VPS:

cd deploy
cp .env.example .env
docker compose up -d --build

MCP endpoint: http://<host>/wiki-mcp/mcp with header X-Api-Key: <MCP_API_KEY>.

See deploy/README.md for details.

Configuration

wiki.config.yaml in your wiki root:

types:
  - entities
  - concepts
  - sources
  - synthesis
  - topics
git:
  enabled: true    # false = local writes only, no commit/push
  branch: main
  commit_prefix: "wiki:"
search:
  refresh_interval_ms: 300000

Roadmap

See ROADMAP.md for planned features (local image assets, rich query outputs, search upgrades).

Why this works

RAG rediscovers knowledge on every question. An LLM wiki compiles once and stays current — cross-references exist, contradictions are flagged, synthesis reflects everything you've read. The MCP is the safe write interface: single-writer lock, index/log maintenance, structured frontmatter.

Browse the wiki in Obsidian; let the agent maintain it.

License

MIT