OKF stands for Open Knowledge Format, an open specification published by Google Cloud that represents organizational knowledge as a directory of plain markdown files carrying YAML frontmatter, a human-readable block of structured metadata. Each file describes one concept: a database table, an application programming interface endpoint, a business metric, a playbook. The format exists because AI agents cannot answer questions about a company's own systems when the answers sit scattered across wikis, code comments, catalog interfaces and the memory of whoever built the pipeline. OKF standardizes the container knowledge travels in, not the knowledge.

What a bundle contains

The unit of distribution is a bundle, a directory tree shipped as a git repository, an archive, or a subdirectory of a larger one. A concept is exactly one markdown file, and its identity is its path with the .md suffix removed, so tables/orders.md carries the identifier tables/orders. There is no registry and no minted identifier scheme; the filesystem does the naming. Two filenames are reserved: index.md lists a directory's contents for progressive disclosure, and log.md records changes under ISO 8601 date headings, newest first.

Each file has two parts: a frontmatter block delimited by three hyphens on their own line, and a markdown body. Exactly one field is required: type, a free-text string with BigQuery Table, API Endpoint, Metric and Playbook offered as examples. Four more are recommended: title, description, resource (a uniform resource identifier for the underlying asset) and tags.

Concepts link to one another with ordinary markdown links, the recommended form resolving against the bundle root so it survives a move. A link asserts that two concepts are related. It does not say how. The relationship lives in the surrounding prose, which a consuming agent has to read every time.

Conformance is permissive. A bundle conforms if every non-reserved file parses as frontmatter and carries a non-empty type. Type values are registered nowhere, and consumers must not reject a bundle over unfamiliar types, missing fields, broken links or absent indexes.

The trust signals added in version 0.2

Version 0.2 arrived on July 24, 2026, six weeks after the first release. Once agents write ten thousand concepts overnight, the implicit accountability of a human-authored page disappears, and a consumer, often another agent, judges each document on explicit signals.

The sources field records the materials a concept derives from, each entry naming a followable artifact or a scope descriptor. Entries may carry three credibility signals, author, usage_count and last_modified, with a usage_window sibling framing counts as a date range. Google declined to store a credibility score at all, on the grounds that scores are subjective and go stale.

The generated field records who produced the content and when it last meaningfully changed; verified is a list of confirmation events. Both use one actor convention: producer and version for agents, a human prefix for people, a process prefix for jobs. From verified a consumer derives a trust tier: unverified when absent, machine-confirmed when only machine actors appear, human-reviewed when a human-prefixed actor does. Tiers are advisory, not access control.

The status field moves a concept through draft, stable and deprecated, absence meaning stable. The stale_after field is a single absolute date rather than a relative expiry, reducing staleness to a plain date comparison. Two renames break compatibility with version 0.1: timestamp is superseded by generated.at, and a body Citations list by frontmatter sources. Consumers may fall back to either form, so a version 0.1 bundle drops in unchanged.

Attested computations

The fifth signal is a concept type rather than a field. An Attested Computation binds a value to a sanctioned way of producing it. The failure mode named is blunt: an agent reporting a currency figure after improvising its own query rather than running the approved one.

The contract sits in frontmatter. The runtime field is required for this type and names how the computation runs, with bigquery, postgres, dbt, python and Looker given as examples, and it fixes what a parameter means, since the word denotes a bind variable in one runtime and a function argument in another. The parameters field lists typed named holes, each carrying name, type and required; an agent may supply values for them and nothing else, and must never edit the computation. The executor field names run instructions and declares the receipt a run must return: a job identifier, the query text actually executed, the result. The attester field names deterministic code, explicitly without a language model, that inspects that receipt and returns a verdict.

Neither receipt nor verdict is written back into the bundle. Verified confirms a definition still matches policy: document-level, slow, stored. Attestation confirms one run produced one value correctly: per-call, discarded. A stale definition can still attest cleanly, and a freshly verified one still needs attestation on every run.

Origin and evolution

In April 2026 Andrej Karpathy published a gist arguing that an agent should incrementally build a persistent cross-linked markdown wiki rather than re-derive knowledge from raw documents on every query. The pattern was already being improvised in AGENTS.md and CLAUDE.md convention files, in Obsidian vaults wired to coding agents, and in metadata-as-code repositories. None of it was designed to interoperate.

Google published version 0.1 on June 12, 2026 in the Data Analytics section of the Google Cloud Blog, authored by Sam McVeety, tech lead for data analytics, and Amir Hormati, tech lead for BigQuery. The specification ran 451 lines and 14.7 kilobytes. Two reference implementations shipped with it: an enrichment agent that walks a BigQuery dataset and drafts a concept per table, and a static visualizer that renders a bundle as a graph. Knowledge Catalog, formerly Dataplex, was updated to ingest OKF that day.

Version 0.2 more than doubled the document. As of August 26, 2026 the specification runs 1,003 lines and 36.7 kilobytes, and the repository shows roughly 8,600 stars, 736 forks, 98 open issues and 76 open pull requests, against two pull requests and 12 issues at launch.

Why it reaches advertising

Nothing in OKF is an advertising product. The link is context governance, which advertising reached from another direction. IAB Tech Lab named its umbrella agentic initiative AAMP on February 26, 2026, covering execution protocols, interoperability standards and an agent registry. AAMP 2.3 added a pricing provenance field to stop agents inventing numbers they had no source for, the bid-request version of what an attester does inside a warehouse.

Cost is the second. Draft Digital collapsed twelve Model Context Protocol calls into a single buyer agent to contain token consumption, and agents on one routing network consumed close to five times as many tokens as human users. Frontmatter a consumer can filter on before opening a document answers that arithmetic directly.

Limitations and disputes

The untyped link is the central design bet and the loudest objection. OKF delivers structural interoperability, not semantic interoperability: two conformant bundles may type the same thing Metric in one and KPI Definition in the other, with nothing telling a machine the difference. Gianluca Fiorelli, writing on July 13, 2026, credited Marc Bara with that distinction and set OKF against the semantic web stack, where relationships carry typed predicates. The trade is deliberate, near-zero authoring cost against precision.

Governance is the second gap. OKF carries an Apache 2.0 license but no standards body home, unlike Model Context Protocol and A2A, the agent-to-agent protocol, both under the Linux Foundation. Fiorelli's formulation was open license, not yet open governance. He also noted that the specification requires one field while Google's own reference parser is stricter.

The format carries no security machinery: no authentication, no authorization, no provenance signing. A bundle is prose an agent reads and acts on, and permissive conformance obliges consumers to tolerate content they did not author. That is an indirect prompt-injection surface by construction.

Discovery is unresolved on the public web. Nothing goes looking for a bundle on a website, and Search Central guidance from May 15, 2026 told site owners that markdown page variants are not used by Search. John Mueller and Microsoft's Fabrice Canel had already warned against crawler-specific markdown in February 2026.

Disambiguation

Open Knowledge Foundation. The same three letters belong to a London nonprofit founded by Rufus Pollock on May 20, 2004, focused on open data and open access. The collision is live: the metadata description on Google's own version 0.2 post calls the specification an Open Knowledge Foundation spec.

Model Context Protocol. MCP governs how an agent reaches tools and data sources; OKF describes what the agent knows. An MCP server can expose a bundle as a source, so the two complement each other.

Knowledge Catalog. A Google Cloud product, formerly Dataplex, that ingests bundles. OKF is the format; Knowledge Catalog is one runtime consuming it.

llms.txt. A community proposal for a curated markdown index served on a public website. OKF targets internal corpora.

Recent developments

Google Cloud documented how to move bundles out of git and into Knowledge Catalog in a post carrying an August 27, 2026 dateline that circulated on August 26. Setup registers an entry group, an okf-bundle entry type and an okf aspect type of thirteen fields. A nine-concept sample bundle produces seventeen entries. Six scalar fields drive server-side search predicates; three array fields, sources and verified among them, do not, forcing client-side filtering after a full record fetch.

That sharpened the vendor-neutrality question. Marie Haynes concluded on August 26, 2026 that GitHub remains her preference for most uses. A specification introduced as requiring no proprietary runtime now has its most detailed scaling path documented for one vendor's catalog alone, with no pricing or adoption figures published.

Timeline

  • April 2026 - Andrej Karpathy publishes a gist describing the LLM wiki pattern, arguing agents should maintain a persistent cross-linked markdown corpus
  • June 11, 2026 - Amir Hormati commits the reference enrichment agent to the GoogleCloudPlatform/knowledge-catalog repository
  • June 12, 2026 - Google Cloud publishes OKF version 0.1, a 451-line specification with one required field, alongside two reference implementations and three sample bundles; Knowledge Catalog is updated to ingest bundles
  • June 17, 2026 - Google announces the Agentic Resource Discovery specification, whose catalog entries can advertise OKF bundles
  • July 13, 2026 - Gianluca Fiorelli publishes an analysis arguing OKF has an open license but not open governance
  • July 24, 2026 - OKF version 0.2 adds provenance, trust, lifecycle and attestation families plus the Attested Computation concept type
  • August 26, 2026 - Google Cloud documentation on pushing bundles into Knowledge Catalog begins circulating, carrying an August 27 dateline
  • December 31, 2026 - stale_after value carried by the sample revenue metric in the acme_retail bundle

Summary

Who: Google Cloud publishes and controls the specification, authored by Sam McVeety, tech lead for data analytics, and Amir Hormati, tech lead for BigQuery. Producers include people, agents and export pipelines; consumers include agents, search indexes, visualizers and deterministic code. Independent practitioners including Gianluca Fiorelli and Marie Haynes have supplied the sharpest published assessments.

What: An open specification representing organizational knowledge as a directory of markdown files with YAML frontmatter. One field, type, is required. Version 0.2 adds optional provenance, trust, lifecycle and attestation families plus the Attested Computation concept type, which binds a value to a sanctioned computation and a deterministic check.

When: Version 0.1 was published on June 12, 2026, and version 0.2 on July 24, 2026. Documentation for pushing bundles into Knowledge Catalog circulated on August 26, 2026.

Where: The specification, reference implementations and sample bundles sit in the GoogleCloudPlatform/knowledge-catalog repository on GitHub under an Apache 2.0 license. Bundles may live in any git repository, on any filesystem, or inside Google Cloud's Knowledge Catalog.

Why: Agents that cannot find authoritative internal definitions produce wrong answers about metrics, schemas and processes. Every agent builder currently solves that context-assembly problem alone, and every catalog vendor rebuilds the same data model behind a proprietary interface. OKF proposes a common file-level container any producer can write and any consumer can read, with enough frontmatter signal for a consumer to judge trust before spending tokens on prose.