LLM Wiki

← Docs hub

FAQ

Is my data sent anywhere?

No. Everything runs locally. There is no telemetry, no analytics, no phoning home. The only network request the built site makes is loading highlight.js from a CDN for syntax highlighting, and that degrades gracefully if blocked.

The --synthesize flag on llmwiki build calls the local claude binary on your machine for overview generation. It is opt-in and does not use any external API directly.

Which agents are supported?

Agent Registry name Status
Claude Code claude_code Production
Codex CLI codex_cli Production
GitHub Copilot Chat copilot_chat Production
GitHub Copilot CLI copilot_cli Production
Cursor cursor Scaffold (SQLite parser in progress)
Gemini CLI gemini_cli Scaffold (schema TBC)
Obsidian obsidian Production (vault as input source)

The PDF adapter that used to live here was removed in the simplification sweep.

See multi-agent-setup.md for per-agent details.

Do I need an API key?

No. llmwiki reads existing session files from your agents' local session stores. It does not call any LLM API. The wiki layer (/wiki-ingest, /wiki-query) uses your existing Claude Code (or other agent) session to do the LLM work.

Can I use it offline?

Yes. The entire build pipeline runs offline. The only external resource is highlight.js loaded from a CDN in the generated HTML for syntax highlighting. If the CDN is unreachable (firewall, offline), code blocks render as plain text -- everything else works normally.

How do I add a new adapter?

See adapter-authoring.md. The short version: extend BaseAdapter, set session_store_path, implement is_available(), add the @register decorator, and import it in discover_adapters().

Where is my data stored?

raw/       Converted session transcripts (markdown)
wiki/      LLM-maintained wiki pages
site/      Generated static HTML

All three directories are gitignored by default. They never enter version control unless you explicitly un-ignore them.

The converter state file (.llmwiki-state.json) tracks which sessions have been processed and is also gitignored.

Can I deploy to GitLab Pages?

Yes. See deploy/gitlab-pages.md. Copy .gitlab-ci.yml.example to .gitlab-ci.yml, push, and the pipeline builds and deploys automatically.

For GitHub Pages, see deploy/github-pages.md.

What's the wiki layer for?

The wiki layer implements layer 2 of Karpathy's LLM Wiki spec. While raw/ holds immutable session transcripts and site/ holds generated HTML, wiki/ is maintained by your coding agent:

Pages interlink via [[wikilinks]] and compound over time. The wiki layer is optional; the site builds from raw/ alone.

The built site includes a client-side search powered by a pre-built search-index.json. Access it via:

Search covers session titles, project names, and content snippets.

Can multiple agents write to the same wiki?

Yes. Each agent produces sessions with its own adapter name and project slug, so there are no collisions in raw/sessions/. The wiki layer (wiki/) is agent-agnostic -- entities and concepts from different agents merge naturally.

For example, a Claude Code session about "React hooks" and a Copilot Chat session about "React hooks" both contribute to the same wiki/entities/React.md page.

How do I update a single wiki page without re-ingesting everything?

Use the /wiki-update command inside your coding agent session. It performs a surgical edit of one page without running the full ingest workflow.

What Python version do I need?

Python 3.9 or later. The only runtime dependency is the markdown package. Everything else uses the standard library.

Can I export to other formats?

Yes. llmwiki supports several export formats:

Run llmwiki export all to generate everything at once.

The export-obsidian, export-qmd, export-marp, and watch subcommands were removed in v1.2.0. See docs/UPGRADING.md for the migration path.