llmwiki

Compile knowledge once, maintain it continuously.

A portable, headless LLM Wiki CLI. An LLM incrementally builds and maintains a persistent, interlinked markdown knowledge base from your sources — so knowledge compounds instead of being re-derived on every query (RAG).

This is Andrej Karpathy's LLM Wiki pattern, as a reusable engine. Bring your own key (any OpenAI-compatible endpoint — OpenAI, DeepSeek, Ollama, OpenRouter — or Anthropic).

Install

npm install -g llm-wiki-agent

Quickstart

llmwiki init                              # scaffold a knowledge base
llmwiki ingest ./paper.md                 # two-step ingest → compiled, cross-linked pages
llmwiki ask "what are the main claims?"   # retrieval + synthesis, cited [[like-this]]
llmwiki lint --fix                        # deterministic hygiene (relocate/normalize/repair)
llmwiki maintain                          # fill knowledge gaps, refresh stale pages
llmwiki health                            # is the wiki compounding? (coverage/citations/freshness/…)

With a model provider:

llmwiki --provider openai --base-url https://api.deepseek.com/v1 \
        --model deepseek-chat --api-key $KEY ingest ./paper.md

The deterministic commands (init, list, read, search, graph, lint, health, insights, impact, index rebuild) work with no model at all. Use --mock to exercise the full ingest→ask loop with a scripted LLM (no key).

What's special

• Owned autonomy — the engine plans + runs maintenance itself (no external scheduler); --auto gates on accumulated signal.
• Quality evals — health makes "is the wiki compounding?" measurable.
• Hybrid retrieval by default — BM25 + optional vectors (RRF) + graph expansion; CJK content is first-class.

Related packages

• llmwiki-core — the headless engine.
• llm-wiki-agent-mcp — MCP server (drive from Claude Code / Codex).
• llmwiki-skill — portable Skill bundle.

Repo + architecture: see the project README.

MIT.