okf-kit

Turn any website into a portable, agent-ready knowledge bundle — no LLM required to start.

okf-kit crawls a site into a Google Open Knowledge Format (OKF) bundle: a directory of markdown concept files with YAML frontmatter and per-directory index.md listings that any agent can navigate with plain file reads. Build it, keep it in sync as the site changes, publish it, and chat with it — locally, with your own key, or fully offline via Ollama.

pip install okf-kit
okf build https://docs.example.com -o docs-okf   # crawl → OKF bundle (no key, no browser)
okf chat docs-okf --provider ollama              # chat offline, no key

Or zero-install with uv:

uvx --from okf-kit okf build https://docs.example.com -o docs-okf

Part of the calknowledge ecosystem — okf-kit is the lightweight, open library; calknowledge is the full platform (LLM enrichment, RAG export, retrieval evals, GUI) built on top of it.

Why

Everyone re-crawls and re-indexes the same docs privately and badly. okf-kit makes a website's knowledge a portable artifact:

• Agents can read an OKF bundle; they can't read your website. The bundle is navigable markdown — no scraping, no SDK, no runtime.

• Faithful markdown, not text soup. Real extraction (headings, code, tables), boilerplate filtered, JS-rendered when needed.

• Self-maintaining. okf sync updates only what changed, so a published bundle in git produces small delta commits and never goes stale.

• Works with any LLM, or none. Chat via OpenAI, Ollama, vLLM, OpenRouter, or Claude — or get a zero-key retrieval answer with citations.

Related MCP server: Crawl4AI RAG MCP Server

Install

pip install okf-kit                 # core: build / sync / validate / zip / list / get / visualize
pip install "okf-kit[chat]"         # okf chat via OpenAI-compatible providers (OpenAI, Ollama, …)
pip install "okf-kit[anthropic]"    # Claude as a chat provider
pip install "okf-kit[js]"           # crawl JavaScript-rendered sites (pulls a Playwright Chromium)
pip install "okf-kit[mcp]"          # serve bundles to Claude Code / Cursor over MCP
pip install "okf-kit[enrich]"       # okf build --enrich (LLM descriptions + tags)

The default install has no browser and no LLM SDK — it installs in seconds.

Tip: install into a dedicated virtualenv so okf-kit's dependencies don't mix with your other projects:

python3 -m venv ~/okf && ~/okf/bin/pip install okf-kit

This also avoids clashes if an existing environment already pins packages like lxml (e.g. a prior crawl4ai install) — a plain install would otherwise bump them.

Commands

Build

okf build https://docs.example.com -o docs-okf --max-depth 3 --max-pages 200

Domain-restricted BFS crawl → an OKF bundle: pages/ mirror with frontmatter concepts, a .okf-kit/state.json for sync, and an index.md in every directory for agent navigation. Validated on exit. No API key needed.

By default the crawl is scoped to the seed's path section — okf build https://doc.rust-lang.org/book/ stays under /book/ and won't wander into the rest of the host. Override with --path-prefix PATH (a narrower/different scope) or --all-paths (the whole host). Other flags: --js (JS-rendered sites — build hints when a site needs it), --no-robots, --enrich (add LLM descriptions/tags — needs [enrich] + OPENAI_API_KEY).

Sync

okf sync docs-okf

Re-crawls the same site and updates only the delta — added pages written, changed pages rewritten, removed pages deleted, unchanged pages left byte-for-byte (stable git diffs). A safety valve aborts on a suspiciously empty re-crawl (--force overrides).

Chat

okf chat docs-okf --provider ollama                 # offline, no key
okf chat docs-okf --provider openai --trace         # any provider, with citations + a navigation trace
okf chat docs-okf                                   # no provider → zero-key retrieval answer
okf chat docs-okf --resume                          # continue the last session (history is local)

The agent navigates the bundle (list_directory / read_concept) to the most specific concept and answers only from what it read, citing the paths.

Chat history is stored locally at ~/.okf/chats/<bundle>/.

Visualize

okf visualize docs-okf          # -> docs-okf/graph.html

A self-contained interactive graph (nodes = concepts, edges = internal links); no backend, no CDN — open the HTML from file://.

Serve over MCP

okf serve-mcp docs-okf          # or --all for every downloaded bundle

Exposes list_bundles / list_directory / read_concept / search_bundle over stdio MCP for Claude Code/Desktop, Cursor, and any MCP client.

Or run it as a container (the included Dockerfile bakes in the rust-book bundle):

docker build -t okf-kit-mcp .
docker run -i --rm okf-kit-mcp   # speaks MCP over stdio; serve another bundle: … okf-kit-mcp okf serve-mcp <name>

Serve a local API (for GUIs)

pip install "okf-kit[serve]"
okf serve                        # prints {"event":"ready","url":…,"token":…}

A loopback-only HTTP API that wraps registry / read / chat / settings, so a desktop app or web UI can be pure UI over an API (no duplicated logic). Guarded by a per-launch bearer token. Endpoints cover browsing the registry, installing/ removing books, reading (toc + concept with heading anchors), chat with saved sessions and cited, deep-linkable sources, and settings (API key kept in the OS keychain). Consume-only, so it stays light to bundle.

Built on this: okf-desktop — a lightweight desktop app (React + pywebview) that reads and chats with your bundles like a book, entirely over okf serve.

Registry

okf list --remote               # browse published bundles
okf get backstage-docs          # download, validate, install to ~/.okf/bundles/
okf list                        # your local bundles

Package for hand-off

okf zip docs-okf                # -> docs-okf.zip, ready to publish or share

Publishing

See docs/PUBLISHING.md — build a bundle, ship it as a release zip with a weekly self-sync Action, and add it to the awesome-okf-kit registry. Publish only content you may redistribute.

Bundle layout

docs-okf/
    index.md                 root directory listing (reserved, no frontmatter)
    log.md                   build/sync history
    pages/                   one concept per page (frontmatter + body + citations)
        index.md             directory listing (every directory has one)
        home.md
        docs/…
    .okf-kit/state.json      crawl config, per-page content hashes, link edges

FAQ

Does okf-kit require an LLM or API key? No. The entire build path — crawl, structure, validate — runs with zero API keys and zero model calls. An LLM is optional: you only need one for synthesized okf chat answers (use Ollama for fully offline) or the optional --enrich step. With no model configured, okf chat still answers from a zero-key retrieval fallback, with citations.

What is OKF, and is okf-kit official? OKF (Open Knowledge Format) is an open, vendor-neutral spec for representing knowledge as markdown files with a little YAML frontmatter (type, title, description, resource, tags, timestamp), introduced by Google as part of its Knowledge Catalog work. okf-kit is an independent, unofficial implementation — and it interoperates: it validates and renders Google's own reference bundles unchanged.

How is okf-kit different from Google's OKF tools? Google's reference implementation targets BigQuery / data-catalog metadata with an LLM enrichment agent. okf-kit targets any website with a deterministic crawler that needs no LLM, no cloud, and no API key — and adds incremental sync, offline chat, an MCP server, and a community registry. Same format, complementary mission.

Can I use it with Claude Code / Cursor? Yes — okf serve-mcp <bundle> is a stdio MCP server exposing list_bundles, list_directory, read_concept, and search_bundle, so any MCP client can read a project's up-to-date docs locally. There's also a Docker image.

Does it work offline? Yes. Building a bundle needs network only to crawl the site once; after that it's all local files. okf chat runs fully offline with Ollama, or with the zero-key retrieval fallback.

What sites can it crawl? Any static or server-rendered HTML site (docs, wikis, blogs). It respects robots.txt and scopes the crawl to the seed's path by default. JavaScript-rendered SPAs need the optional [js] extra (a real browser); extraction quality varies by site.

Development

pip install -e ".[dev]", then pytest -q (fully offline) and ruff check okf_kit tests. See CONTRIBUTING.md and the CHANGELOG.

License

Apache-2.0.