Google Cloud published the Open Knowledge Format on June 12, a vendor-neutral spec that represents the context AI agents need as a directory of plain markdown files with YAML frontmatter. Tech leads Sam McVeety and Amir Hormati wrote it. It runs about 451 lines and demands exactly one field from every document. And on the same day it shipped, Google wired the format into the Knowledge Catalog product it charges customers to run.
Hold that last part. We'll come back to it.
So what is it, actually
An OKF bundle is a folder. Each concept, whether that is a database table, a metric definition, a runbook, or an API, lives in its own markdown file. Files link to each other with ordinary markdown links, and those links form the knowledge graph. The only mandatory frontmatter field is type. Everything else (title, description, resource, tags, timestamp) is recommended and optional. There are two reserved filenames, index.md and log.md, and the spec is blunt about its own simplicity. According to the GitHub repo, if you can cat a file you can read OKF, and if you can git clone a repo you can ship it.
That is the whole pitch on the surface. No SDK. No runtime. No database. The company blog frames the value as coming from how many parties speak the format rather than who owns it, which is the kind of line you write when you want a thing adopted, not bought.
Google shipped reference tooling too: an enrichment agent that walks a BigQuery dataset and drafts concept documents, a static HTML visualizer that renders the bundle as an interactive graph, and three sample bundles. Released under Apache 2.0.
The pattern Google didn't invent
Here's the thing. None of this shape is new. OKF formalizes what Andrej Karpathy calls the LLM wiki, a folder of markdown an agent reads and maintains on its own. The gist has thousands of stars. The same idea has been spreading for a while under other names: Obsidian vaults wired to coding agents, the AGENTS.md and CLAUDE.md files developers drop into their repos, assorted "metadata as code" experiments. AGENTS.md alone is used by more than 60,000 open-source projects, now stewarded by the Agentic AI Foundation under the Linux Foundation.
So conforming to OKF costs almost nothing, because half the tooling already does something close to it by accident. That's exactly why Google had reason to name it. A convention people already follow spreads without a sales team behind it.
"LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass," Karpathy wrote in the gist Google cites, which is a genuinely good argument for why machines can maintain a wiki that humans always abandon.
What none of the earlier versions had was a vendor with a catalog to serve them from. And that gap is the whole story.
Where the money actually is
The spec is precise about what it refuses to do. It explicitly lists storage, serving, and query infrastructure as non-goals. It does not handle access control. A markdown file can tell you what a table means. It cannot decide which agent is allowed to read it, log that access, or stop the data from walking out the door.
Those omissions map almost exactly onto Google's paid layer. Storage, serving, query, access control: that is what the Knowledge Catalog does, the product Google detailed earlier this year and rebranded from the old Dataplex line. Hormati wrote that OKF carries no dependency on any cloud, model, agent framework, or catalog. In the same post he noted Google had updated its Knowledge Catalog to ingest OKF natively and serve it to agents. The vendor-neutral claim and the product pitch sit one sentence apart, which, frankly, tells you most of what you need to know about the strategy.
Michael Gerstenhaber, Google Cloud's VP of product management for its agent platform, made the value chain explicit in a June 11 interview with Computer Weekly. He framed the hard problem as access, not authorship: he'd only get comfortable giving a virtual employee access to sensitive data through identity, permissioning, audit, and observability. None of that is in OKF. All of it is in the layer above.
Give away the file. Meter the gateway.
Does it become a standard or not?
This is the part I can't call yet. A format is only interoperable if producers outside Google adopt it, and right now every sample bundle in the repo was built by Google. Whether the metadata catalog vendors (Atlan, Alation, Collate, the usual suspects) decide to emit OKF is the entire question. If they do, Google gets credited with the lingua franca and sells the serving layer underneath it. If they don't, OKF is a tidy folder convention with a Google logo on it.
One more wrinkle worth noting: there's a separate, unrelated supply-chain spec also called OKF floating around, so the name isn't even clean.
v0.2 is supposed to take community feedback. The spec mostly nails structural interoperability, the folder layout and the required field, and punts on semantic interoperability, what the concepts actually mean across organizations, to "future conventions." That's the hard part, and it's the part still unwritten.
