Google Announces Open Knowledge Format (OKF) for AI Agent Knowledge

This article was generated by AI. The research and factual claims are based on publicly available sources linked within.

Google Cloud has published the Open Knowledge Format (OKF), a vendor-neutral specification that standardises how organisations represent knowledge as portable markdown files. The spec — version 0.1, released on June 12, 2026 — formalises the "LLM wiki" pattern popularised by Andrej Karpathy into an open, interoperable format that any AI agent can read without a proprietary SDK or database.

Authored by Google Cloud tech leads Sam McVeety and Amir Hormati, the specification lives on GitHub under the Apache 2.0 licence in the GoogleCloudPlatform/knowledge-catalog repository, where it has already gathered over 3,700 stars.

The problem OKF solves

Organisational knowledge is fragmented. It lives across metadata catalogs with proprietary APIs, wikis, shared drives, code comments, docstrings, and the heads of senior engineers. Every team building an AI agent solves the same context-assembly problem from scratch, and every catalog vendor reinvents the same data models. Knowledge ends up locked inside whichever system created it.

OKF addresses this by defining a common surface — a "lingua franca" — that any producer or consumer can speak. The file is the contract. Tooling on each end is freely swappable.

Three design principles

Minimally opinionated. Only a type field in YAML frontmatter is required. Everything else — title, description, tags, timestamps, resource links — is up to the producer. The spec defines the interoperability surface, not the content model.

Producer/consumer independence. A human can handwrite a bundle that an AI agent consumes. An automated pipeline can generate a bundle that a human browses in a visualiser. The format is the contract, and the tooling on each end is decoupled.

Format, not platform. OKF is never tied to a specific cloud, database, model provider, or agent framework. No proprietary account or SDK is required to read or write it.

Bundle structure

An OKF bundle is simply a directory of markdown files with YAML frontmatter:

my_bundle/
├── index.md
├── log.md
├── datasets/
│   ├── index.md
│   └── sales.md
└── tables/
    ├── index.md
    ├── orders.md
    └── customers.md

Each concept document contains frontmatter with at minimum a type field, plus optional fields like title, description, resource, tags, and timestamp. The markdown body can include schema tables, join paths, cross-links to other concepts, and any additional prose. Concepts link to each other via standard markdown links, turning the directory into a navigable knowledge graph.

Three conformance rules make a bundle valid: every non-reserved .md file must contain parseable YAML frontmatter, every frontmatter block must contain a non-empty type field, and reserved filenames (index.md, log.md) must follow their defined structure.

Reference tools

Google shipped three complementary tools alongside the spec:

• BigQuery Enrichment Agent — Walks a BigQuery dataset and drafts an OKF concept document for every table and view, with a second LLM pass that crawls authoritative documentation for enriched schemas and join paths.
• Static HTML Visualizer — Turns any OKF bundle into a self-contained interactive HTML file with a force-directed graph, detail panel, search box, type filter, and "cited by" backlinks. No backend required.
• Three sample bundles — Based on GA4 e-commerce, Stack Overflow, and Bitcoin public datasets.

Adoption and community response

The GitHub repository has seen strong early engagement, and third-party tooling has already emerged — including a Rust okf crate, an open-source conformance validator from WitsCode, and community documentation at openknowledgeformat.com.

As a v0.1 draft, the spec is explicitly a starting point. Google has invited other vendors and the community to adopt, extend, and build their own implementations. Whether OKF becomes a true industry standard will depend on adoption beyond Google — from platforms like Atlan, Alation, and Collibra — but the barrier to entry could hardly be lower.

"If you can cat a file, you can read OKF; if you can git clone a repo, you can ship it." — OKF Specification v0.1

The full specification and all reference code are available at github.com/GoogleCloudPlatform/knowledge-catalog.