Awesome OKF

A curated list of resources for the Open Knowledge Format (OKF) — Google's open, vendor-neutral spec for representing knowledge as Markdown files with YAML frontmatter that AI agents can read without custom integrations. This list practices what it catalogs: it is itself published as a conformant OKF bundle (see below).

OKF was announced by Google Cloud in June 2026; v0.2 followed in July 2026. It formalizes the "LLM-wiki" pattern into a portable, interoperable format: knowledge lives as plain Markdown files with structured frontmatter (type, title, description, resource, tags, plus optional provenance, trust, and lifecycle fields) and cross-links that form a knowledge graph. No SDK, no runtime, no lock-in — just files you can ship in a tarball, host in Git, or mount on a filesystem.

Contents

• OKF at a Glance
• Specification
• Official Tools & Reference Implementations
• Sample Bundles
• Community Tools
• Articles & Guides
• Background & Origins
• Built on the LLM-Wiki Pattern
• Related Formats & Concepts
• Community

OKF at a Glance

A bundle is a directory tree of Markdown files, and the directory structure is independent of the domain. The essentials from the v0.2 spec:

Aspect	Rule
Concept	One Markdown file: YAML frontmatter plus a free-form Markdown body.
Required field	type — a non-empty string identifying the kind of concept.
Recommended fields	title, description, resource (a URI for the underlying asset), tags.
Provenance & trust	Optional sources (with per-source credibility signals), generated ({ by, at }; supersedes v0.1 timestamp), and verified — consumers derive trust tiers (unverified / machine-confirmed / human-reviewed).
Lifecycle	Optional status (draft / stable / deprecated; absent means stable) and stale_after (an absolute date).
Attested computations	A type: Attested Computation concept carries a sanctioned computation (runtime, parameters, executor, attester) so a consumer can confirm a value was produced the blessed way, not improvised.
Actors	generated.by / verified[].by use <producer>/<version> for agents, human:<id> for people, process:<id> for processes.
Reserved filenames	index.md (directory listing for progressive disclosure) and log.md (update history); every other .md is a concept. A bundle-root index.md may declare okf_version in frontmatter.
Links	A link from concept A to B asserts a relationship; use bundle-relative paths (starting with /) or ordinary relative paths.
Conformance	Every non-reserved .md file has parseable frontmatter with a non-empty type; reserved files follow their defined structure when present.
Consumer tolerance	Consumers must tolerate missing optional fields (including every trust and provenance family), unknown type values, broken links, and a missing index.md.
Extensions	Producers may add custom keys; consumers must preserve unknown fields.

This list is also published as a conformant OKF v0.2 bundle under bundle/, generated from this README by scripts/build-okf-bundle.mjs.

Specification

• OKF v0.2 Specification (SPEC.md) - The universal, vendor-neutral spec, including conformance criteria and reserved filenames. v0.2 (July 2026) makes provenance (sources), trust (generated / verified), lifecycle (status / stale_after), and attested computations first-class; it supersedes v0.1's timestamp field and body # Citations list. The v0.1 text stays readable at a pinned commit.
• GoogleCloudPlatform/knowledge-catalog okf/ - The home repository for the format, reference code, and samples.
• OKF README - Overview of OKF philosophy, installation, and usage.

Official Tools & Reference Implementations

• Reference Agent - Python implementation (built on Google's Agent Development Kit) that produces and visualizes OKF bundles: it drafts OKF documents from a BigQuery source, enriches them with web-crawled provenance (sources frontmatter with per-claim footnote attribution, since v0.2), and ships a viewer/generator.py that renders the bundle as a self-contained, interactive graph.
• Knowledge Catalog Enrichment toolbox - A ready-to-use agent and customizable harness (TypeScript) to produce, evolve, and maintain metadata in Knowledge Catalog. Includes a server that exposes a Markdown fileset as an MCP server.
• Knowledge Catalog mdcode - Manage metadata as source-code artifacts, with git-style pull/push sync between OKF Markdown and BigQuery / Dataplex / Knowledge Catalog.

Sample Bundles

Four conformant, ready-to-browse bundles from the home repository, each with a self-contained interactive viz.html graph viewer. The first three are built from public BigQuery datasets:

• GA4 e-commerce - Google Analytics 4 e-commerce metadata.
• Stack Overflow - Schema graph for the public Q&A dataset.
• Bitcoin - On-chain concepts from the public Blockchain dataset.
• ACME Retail - The v0.2 showcase: a fictional retailer's finance knowledge exercising Attested Computations (with a deterministic sql_equality.py attester), executor skills, and deprecated lifecycle states.

Community Tools

• okft - Linter and MCP server for OKF bundles (pip install okft): okft lint validates spec conformance and hygiene (broken links, orphans, timestamps) with CI-friendly exit codes and JSON output; okft serve exposes a bundle to any MCP-capable agent as deterministic navigation tools. Apache-2.0.
• BundleDex - Directory of 440+ OKF bundles with search, dedup, categorization, and a JSON API for agents.
• OKF Bundle Generator - Free web tool by Suganthan Mohanadasan: paste a URL or sitemap, it crawls up to 100 pages, converts each into a clean OKF concept, links them into a graph, and outputs a downloadable v0.2 bundle (records generated and sources; verified is left to you).
• SchemaCrawler Scribe - Generates structured database documentation directly from live schema metadata in Google Open Knowledge Format (OKF). See AI-Ready Database Docs You Can Keep in Git for a quick overview.
• samemind - Personal memory for AI coding agents, stored as an OKF bundle: identity, an append-only work-ledger and a kanban board in plain Markdown. export/import speak the OKF v0.1 wire format directly; zero-dep CLI plus an MCP server, with a 12-engine instruction-file installer.
• knowledge-mcp - Public MCP server that exposes an OKF bundle over search, read, list, and related tools; deployed on Cloudflare Workers with no auth. Points at any OKF v0.1 bundle so any MCP client (Claude Code, Cursor, etc.) can query it as a knowledge graph.
• okf-gem - Ruby harness for a bundle's whole life: an agent skill authors and curates concepts, a CLI and library check them (validate for v0.1 conformance, lint as a separate non-blocking curation report, ranked search for retrieval), and okf server / okf render publish an interactive graph, live or as one self-contained HTML file. Runs fully locally; gem install okf or Docker; live demo. Apache-2.0.
• okf-skills - OKF toolkit for Claude Code: a plugin, agent skills, and a GitHub Action to author, validate, and visualize bundles. Built on v0.2 — trust tiers, provenance, staleness, and attested computations. MIT.
• EchoesVault - Persistent-memory plugin for OpenCode that stores its vault as an OKF-compliant bundle — one that doubles as a valid Obsidian vault. MIT.
• claude-mega-brain - Claude Code plugin that injects an OKF knowledge base at session start, with an agentic benchmark against Obsidian-plus-MCP setups. MIT.
• OWOX Model Canvas - Visual, Miro-like editor for data models that reads and writes OKF as its native format; the repo also publishes real bundles. Apache-2.0.

Articles & Guides

• How the Open Knowledge Format can improve data sharing - The official Google Cloud announcement (Sam McVeety & Amir Hormati). Start here.
• OKF v0.2 adds trust signals - The official v0.2 announcement (July 2026): why agent-maintained corpora need provenance, trust, freshness, lifecycle, and attestation answerable straight from frontmatter.
• Introducing the Google Cloud Knowledge Catalog - The companion product OKF was designed alongside; Knowledge Catalog produces and consumes OKF as its open, portable format.
• OKF, by Marie Haynes - Why OKF matters for SEO/AI: the shift from "being found by search engines" to "making knowledge accessible so agents can act on it."
• Build an OKF brain like mine! - Marie Haynes' follow-up: a walkthrough of her personal OKF brain (concepts, entities, playbooks, references, systems), with prompts to build your own.
• Open Knowledge Format (OKF): Google's New Markdown Format for AI Agents - Practical explainer by Suganthan Mohanadasan, updated for v0.2's trust and provenance fields.
• Google shipped an open format (OKF). My site already spoke it. - Hands-on notes from building one of the first community bundles.
• Google Cloud Announces The Open Knowledge Format - Search Engine Journal coverage.
• What Is Google's Open Knowledge Format (OKF)? A Plain-English Guide for Site Owners - Beginner-friendly guide aimed at website owners.
• Open Knowledge Format (OKF): Google AI Agent Standard - Overview of OKF in the context of agentic AI.
• Open Knowledge Format: What Google Actually Shipped - Gianluca Fiorelli's long-form analysis: OKF as enterprise agent memory rather than an SEO play, with early-adopter interviews and the ARD relationship mapped out.
• Composing OKF and ARD on specification.website - Joost de Valk implements both specs on one site, hits real field-name conflicts between them, and files the issues upstream.

Background & Origins

• LLM Wiki, by Andrej Karpathy - The "LLM-wiki" pattern that OKF formalizes. OKF's reserved index.md and log.md files and its raw-sources/wiki/schema layering trace directly back to this gist.
• As We May Think, by Vannevar Bush (1945) - The Memex essay Karpathy cites as the deeper ancestor: a personal, curated knowledge store where the links between documents matter as much as the documents themselves.

Built on the LLM-Wiki Pattern

Community implementations of the LLM-wiki pattern that OKF formalizes. These are not OKF-native, but they share its "knowledge as interlinked Markdown that an agent maintains" philosophy.

• AutoSci - Full-lifecycle AI research platform built on the LLM-wiki vision, with 28 agent skills (slash commands in Claude Code, with Codex and OpenCode adaptations) for ingesting and synthesizing papers.
• karpathy-llm-wiki - Agent Skills-compatible LLM wiki for Claude Code, Cursor, and Codex that ingests sources, compiles wiki pages, answers with citations, and lints for consistency.
• llmwiki - Open-source implementation: upload documents, connect Claude via MCP, and have the agent write and maintain the wiki.
• Synto - Local-first wiki builder using a two-tier Ollama pipeline (small model extracts concepts, larger model writes cross-linked articles) on consumer hardware.
• llm_wiki - Cross-platform desktop app that turns documents into an organized, interlinked knowledge base automatically.

Related Formats & Concepts

• Agentic Resource Discovery (ARD) - Google-announced sister spec, published days after OKF: federated catalogs advertise resources — OKF bundles included — so agents can discover them. OKF packages knowledge; ARD finds it.
• llms.txt - A proposed standard for a Markdown file that helps LLMs use a website's content.
• Model Context Protocol (MCP) - An open protocol for connecting AI agents to tools and data sources.
• AGENTS.md - A simple, open convention for giving coding agents project instructions.
• Obsidian - Markdown vaults with backlinks; a close cousin of the OKF directory-of-notes model.
• Dataview - Query language over Markdown frontmatter, useful for building dynamic views from OKF-style structured fields.
• Marp - Markdown presentation ecosystem; turns Markdown knowledge files into slide decks.

Community

• Google Cloud Tech on X - The launch announcement.
• Google proposes Open Knowledge Format based on Markdown (Hacker News) - The launch discussion: skepticism, prior art, and alternatives.
• awesome-okf (Chinese) - Chinese-language sister list, bundled with Feishu, Obsidian, Notion, and GitHub-to-OKF converters.

Contributing

Contributions are welcome! Please read the contribution guidelines first. The OKF ecosystem is brand new — if you build a producer, consumer, viewer, or bundle, add it here.