LLM Wiki MCP
LLM Wiki MCP gives an agent the LLM Wiki app knowledge base: search, page reading, a link graph and source-grounded chat.
- 21,429 ★ on the repository
- Last commit
Minimal access
Why this level
- The server reads project documents through a local API, and tokenless mode opens access to any local process.
Install
- Works in
Manual install
{
"mcpServers": {
"llm-wiki": {
"command": "node",
"args": ["/absolute/path/to/llm_wiki/mcp-server/dist/src/index.js"],
"env": {
"LLM_WIKI_API_TOKEN": "your-token"
}
}
}
}Build the server first: cd mcp-server, npm install, npm run build.
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The server reads project documents through a local API, and tokenless mode opens access to any local process.
This is third-party code. Review the repository files before installing.
What it does
LLM Wiki, a desktop app, builds and maintains a wiki from your documents following the Karpathy pattern: a model incrementally compiles linked pages instead of searching from scratch for each question. The `mcp-server` folder exposes that app as an MCP server. The server does not read folders itself and calls the app's local API at `127.0.0.1:19828`, so search, permissions and graph match the program window. Tools: `llm_wiki_search`, `llm_wiki_read_file`, `llm_wiki_graph`, `llm_wiki_chat`, `llm_wiki_reviews`, `llm_wiki_rescan_sources`, plus project selection and status.
Who it is for. LLM Wiki users who accumulate documents and notes in the app and want to query them from Claude Code, Cursor or Codex.
Good fit when
- An agent answers from a personal knowledge base rather than model memory.
- Search over accumulated PDFs, articles and notes is needed beside code.
- Gaps and surprising links in the knowledge graph must be found.
- After files land in the sources folder, a rescan is triggered.
Not a fit when
- The LLM Wiki app is not installed and running: the server is useless without it.
- A cloud server without the local app window is needed.
- The knowledge base is kept in plain notes with no wiki building.
Example request
Search my LLM Wiki for notes on RAG, read the two most relevant pages and list related topics from the graph.MCP
- Transport
- stdio
- Authentication
- key is optional
Environment variables
LLM_WIKI_API_TOKENsecret- Local API token of the LLM Wiki app, not needed in tokenless mode
LLM_WIKI_API_BASE_URL- API address when it differs from `http://127.0.0.1:19828/api/v1`
Limitations
It works only while the LLM Wiki app runs with the local HTTP API and MCP access enabled in settings. Node.js 20 or newer is required and the server is built from source with `npm install` and `npm run build`. Pass the API token through the `LLM_WIKI_API_TOKEN` variable, not a command-line argument. Tokenless mode is convenient but opens the API to any process on the machine. Full Deep Research stays in the app. The repository license is not detected automatically.
How to disable. Remove the `llm-wiki` entry from the client MCP settings and switch off Enable MCP access in the app settings.
FAQ
How do I connect the server to a client?
Add a `node` command to the MCP config with the argument `/path/to/llm_wiki/mcp-server/dist/src/index.js` and the `LLM_WIKI_API_TOKEN` variable. Build the server first with `npm run build`.
What must be enabled in the app?
Open settings, the API and MCP section, and enable Enable local HTTP API and Enable MCP access. Then allow tokenless access or set a token.
What does the set_project tool do?
`llm_wiki_set_project` pins the process to one project. Other tools then refuse access to a different project even if the app window switches projects.
Can the server see every project file?
No. Reads go through the API path allow-list and internal app state files are not exposed.
What does llm_wiki_chat do?
It asks the app chat agent a question and returns answer text, source references, token usage and tool events. The `deep` mode widens evidence collection.
Discussion
No questions or reviews yet.
Related
MCP serversAn interactive terminal finder selects files, history commands and entries from long lists by matching parts of their names.
Job search inside your coding agent: scores each listing against your CV from 1 to 5, tailors a PDF resume, drafts letters and tracks applications
A Claude Code workflow where the agent searches job postings, scores fit against your profile, drafts a CV and cover letter, and helps prepare for interviews
A skill that makes the agent open with the action, number the steps and skip tangents
Obsidian skills
Obsidian Skills
Skills for agents working with Obsidian: Markdown with wikilinks, Bases, JSON Canvas, Obsidian CLI, Defuddle and Knap
A self-hosted knowledge base with block-level references and a built-in MCP server for connecting AI agents to your notes
README in short
The MCP server README lists requirements, the build, a client config example, ten tools and the security model: 127.0.0.1 only and a path allow-list.