What OKF is
The format this project compiles, and the state of the ecosystem around it.
The Google Cloud data team published Open Knowledge Format in June 2026 as v0.1. Version 0.2 shipped 25 July 2026.
It is a vendor-neutral open specification for representing organizational knowledge so agents can consume it. In practice it formalizes the LLM-wiki pattern into a portable format: a directory of Markdown files with YAML frontmatter. No SDK, no runtime, no registry.
The problem it targets is context assembly. Internal knowledge lives in metadata catalogs behind proprietary APIs, wikis, shared drives, code comments, and the heads of senior engineers. Every team building an agent solves the gathering problem again from zero.
Structure
Each file is a concept, meaning a table, dataset, metric, API, or runbook.
---
type: BigQuery Table
title: Orders
description: One row per completed customer order.
resource: https://console.cloud.google.com/bigquery?p=acme&d=sales&t=orders
tags: [sales, revenue]
timestamp: 2026-05-28T14:30:00Z
---The body is ordinary Markdown. Links between files form the knowledge graph.
sales/
index.md
datasets/orders_db.md
tables/orders.md
tables/customers.md
metrics/weekly_active_users.mdOnly type is required. Everything else is standardized but optional, and bundles may add their
own keys. index.md gives navigation and log.md gives history. Both optional, and langonrock
skips both, since they are navigation rather than knowledge.
What v0.2 added
Trust and provenance became first-class: sources, generated, verified, status,
stale_after.
If v0.1 asked what your content is, v0.2 asks who stands behind it. That is the part worth
adopting. A plain CLAUDE.md cannot tell you whether it is still true.
Of those, status is the one that reaches the manifest, because it changes whether an agent should
use the concept at all. See the read model for why the others do
not.
Design principles from the spec
The format is minimally opinionated, producers and consumers stay independent, and it is a format rather than a platform. No vendor account required to read it.
That is why compiling it is legitimate. Nothing in the spec asks a consumer to read the files the way the reference implementation does.
What Google actually ships
Repository GoogleCloudPlatform/knowledge-catalog, Apache 2.0.
- The specification,
okf/SPEC.md, currently v0.2 - A reference enrichment agent in Python 3.13. It walks a BigQuery dataset, drafts a concept per table and view, then runs a second LLM pass that crawls documentation and adds citations, schemas and join paths. Only useful if your source is BigQuery.
- A visualizer that renders any bundle as a self-contained HTML force-directed graph. No backend, nothing leaves the page.
kcmd, a CLI and MCP server that syncs with Google Cloud Knowledge Catalog- Four sample bundles: GA4, Stack Overflow, Bitcoin, Acme Retail
Those samples are demos. Google published no general-purpose knowledge corpus.
The ecosystem
BundleDex indexes 381 bundles from 333 authors, of which 148 carry the OKF-conformant badge. Domains cover education, healthcare, infrastructure, finance, COBOL and philosophy. Some come from Databricks and Pulumi, most from individuals.
Quality signals are thin. Stars, a conformance badge, and a Draft marker. There is no unified install path, since each bundle links to its own repository. Treat it as an awesome-list, not a trusted registry.
Community tooling
| Tool | What it does |
|---|---|
okflint | Deterministic conformance checker, Python, MIT |
| Kiso | Validator and static-site builder, Java, Apache-2.0 |
| OpenWiki | Wires codebases into OKF, Python, MIT |
superops-team/okf | Go CLI generating bundles from git repositories |
signed-okf | Cryptographic signatures, optional OriginTrail anchoring |
hermes-okf | Filesystem memory storing agent decisions as concepts |
okf-skill | A single-file agent skill for Claude Code and Cursor |
Obsidian works as-is, since a bundle is already a vault. langonrock keeps that true by never touching your source folder.
LangGraph and LangChain
okf-agents, MIT, Python 3.11 to 3.13.
from okf_agents import OKFBundle, create_okf_tools
bundle = OKFBundle.load("./my_markdown_docs")
tools = create_okf_tools(bundle)It exposes read_concept, search_concepts, list_links, read_index, two retrievers, a router,
and a navigator subgraph with hop and token budgets. That navigator is the baseline the
benchmarks measure against.
Maturity: 4 stars, 28 commits, one person, a few weeks old at the time of writing. Read the code before trusting it. It is small enough to audit in an afternoon or replace outright.
MCP servers
okf-mcp, okf-ingest with verbs context, search, impact, diff and doctor, and
okfbundle.com. All community.
The verdict that motivated this project
There is no ready-made knowledge base to download that makes an agent smarter. What exists is a format plus an uneven community directory.
The real value of OKF today is as a convention for knowledge you write yourself. Your schemas, metrics, runbooks, architecture decisions. That is what this project stores.