xyz-wiki

A wiki your LLM writes and keeps up to date itself, with a typed ontology, works with any agent that supports MCP.

You give the model your stuff and it turns it into small linked pages, one concept, entity or claim per page, with typed relations between them (is_a, part_of, contradicts, ...). Its answers get filed back as pages too, so the wiki keeps growing as you use it, and a lint pass finds what needs fixing. Search is plain keyword search (BM25) plus a walk over the ontology graph, no embedding model, no vector database, no GPU needed.

• The only thing you need is Node 22.13+ (the built-in node:sqlite with FTS5), the code is three files.

• Pages are just markdown files with OKF-style front matter, the index is a cache rebuilt from them, so it all works fine with git.

• It stays quick when it gets big, on our PC a 2,000-page wiki answers a search in under 80 ms and reopens in under a tenth of a second, only the files that changed get read again.

• The MCP server (stdio) works with DeepSeek Harness, Claude Desktop/Code, OpenClaw, Cursor, anything that supports MCP.

• The model does the organizing, the server only stores, indexes and retrieves, and skills/xyz-wiki/SKILL.md tells the agent how to ingest, answer and maintain.

Install

git clone https://github.com/xyznq1/xyz-wiki.git && npm install -g ./xyz-wiki   # Node.js 22.13+, no dependencies
cd xyz-wiki && npm test                                                    # optional: the test suite

Related MCP server: LLM Wiki MCP Server

Connect your agent

The server takes the wiki folder as its argument (or XYZ_WIKI_DIR) and creates it on first use.

DeepSeek Harness: dsh/xyz-wiki.patch.yml is a ready overlay, it runs the agent on your own OpenAI-compatible model server (llama.cpp, vLLM, Ollama) with the xyz-wiki MCP server and the skill, telemetry off. Set the endpoint and model id in the patch, then run:

export XYZ_WIKI_HOME=/path/to/xyz-wiki XYZ_WIKI_DIR=~/xyz-wiki
dsh web --patch $XYZ_WIKI_HOME/dsh/xyz-wiki.patch.yml             # or: dsh headless --patch ... "learn notes.md"

Claude Desktop, Claude Code and Cursor take this MCP config:

{ "mcpServers": { "xyz-wiki": { "command": "xyz-wiki-mcp", "args": ["/path/to/wiki"] } } }

OpenClaw: set mcp.servers.xyz-wiki = { command: "xyz-wiki-mcp", args: ["/path/to/wiki"] }, then allow the tools (xyz-wiki__wiki_search, ...) in tools.allow.

Then give the agent the skill (or paste its steps into the system prompt) and hand it stuff, "learn this", "what do we know about X", "clean up the wiki".

Tools

Vision (optional)

wiki_see shows up when XYZ_WIKI_PYTHON points at a Python that has yolov5 installed. YOLOv5 is AGPL-3.0, so you install it yourself, we don't bundle it. Reading text needs the Tesseract binary on PATH (or TESSERACT), and each half works without the other.

python -m venv ~/xyz-vision && ~/xyz-vision/bin/pip install yolov5
~/xyz-vision/bin/yolo settings sync=False                 # Ultralytics analytics off
export XYZ_WIKI_PYTHON=~/xyz-vision/bin/python XYZ_VISION_WEIGHTS=yolov5s.pt   # fetched on first use

XYZ_VISION_CONF sets the detection threshold (default 0.25). To test it, run XYZ_WIKI_TEST_IMAGE=photo.jpg npm test.

A page

---
title: Fair Value Gap
type: concept
status: draft
aliases:
  - FVG
sources:
  - resource: sources/notes-2026-09.md#fvg
relations:
  - is_a: Price Imbalance
  - related_to: Order Block
generated: 2026-09-28
updated: 2026-09-28
verified: unverified
---
A three-candle pattern where the first and third wicks do not overlap: price moved faster than orders filled.
See [[Order Block]].

type is the only field a reader needs. status (draft/stable/deprecated), sources, verified (unverified/machine-confirmed/human-reviewed) and stale_after follow the Open Knowledge Format (OKF v0.2).

CLI

xyz-wiki ./wiki stats | lint | list [type] | search <words> | read <title> | reindex

Why it's built like this

• No embeddings. A small, well-linked wiki the model wrote itself can be searched by the words it uses, and the ontology covers what keyword search misses (the related block in every search result), so there's no extra service to run and nothing to re-embed when pages change.

• The model writes the pages. Searching raw chunks works the same answers out again every time, a wiki that's kept up to date keeps them (the "LLM wiki" pattern). Answers get filed as type: query pages.

• Tools return snippets and capped page bodies, never the whole thing, so the context cost stays low.

Credits

What's ours: xyz-wiki itself, the MCP server, the index and the ontology search, the lint, the agent skill, the CLI, the DeepSeek Harness overlay and the vision tool.

What we built on, and credit to the people who made it:

• Andrej Karpathy's LLM Wiki idea, the model writing and keeping up its own wiki instead of doing RAG over raw files every time.

• Google Cloud's Open Knowledge Format (OKF v0.2), the front matter the pages use.

• SQLite's FTS5 for the BM25 search, and YOLOv5 by Ultralytics and Tesseract for the optional vision tool.

License

MIT.

For further questions, DM us on Instagram: @xyz_nq1