Google Cloud introduced the Open Knowledge Format (OKF), a new open specification designed to untangle the mess of scattered organizational knowledge. It’s not a product or a platform—just a format. A simple, vendor-neutral Markdown spec with YAML frontmatter that AI agents and humans can both read without special tools.

AI agents choke on context. They need to understand data schemas, metrics, runbooks, and APIs before they can act. Today, that information hides in wikis, metadata catalogs, code comments, and the brains of senior engineers. Each system speaks its own language, forcing every AI developer to reinvent the wheel. OKF cuts through this chaos by standardizing how knowledge is packaged and shared.

At its core, OKF organizes knowledge into directories of Markdown files. Each file represents a single “concept”—a table, a metric, an API endpoint, or a runbook. The file path itself is the concept’s identity. A YAML frontmatter block at the top carries minimal structured data: a required “type” field and optional fields like title, description, resource links, tags, and timestamps. The rest is free-form Markdown. Concepts link to each other with normal Markdown hyperlinks, creating a navigable knowledge graph rather than a rigid database schema.

The format demands almost nothing. Just one required field per file. No new runtimes. No SDKs. No proprietary accounts. The files are human-editable, Git-friendly, and render on GitHub or any Markdown viewer. This simplicity is the point. OKF lives in plain text and leverages the power of existing tools like git for version control, access management, and collaboration. It moves knowledge freely between teams and tools without lock-in.

Google Cloud shipped reference tools alongside OKF’s release. A two-pass enrichment agent crawls BigQuery datasets, drafting OKF documents for each table and view, then enriches these with join paths, business context, and citations by crawling authoritative documentation using a large language model. On the consumer side, a static HTML visualizer renders any OKF bundle into an interactive graph. Both are open source and demonstrate OKF’s practical potential.

Why OKF Matters for AI Teams and Enterprises

Before OKF, every company building AI agents faced the same problem: how to collect scattered knowledge into a form agents can use. Each solved it with custom scripts, siloed wikis, or proprietary metadata catalogs. The result? Fragmented solutions that don’t talk to each other and lock knowledge inside specific tools.

OKF offers a shared language. It’s flexible enough to cover diverse knowledge domains, from datasets to incident playbooks. Its minimal schema means producers and consumers evolve independently. An AI agent can read a bundle created by a team without special adapters. Teams can version their knowledge alongside code, review changes through pull requests, and maintain a single source of truth.

This approach is also a compliance win. Since OKF is just files—no cloud lock-in or SaaS dependency—organizations retain full control over where their knowledge lives. Versioning in git means every change is traceable, crucial for regulated industries and GDPR compliance.

OKF complements existing AI agent protocols. It is not a replacement for runtime standards like the Model Context Protocol (MCP) or retrieval-augmented generation (RAG). MCP handles how agents access live data and tools. RAG searches vast document corpora at inference time. OKF sits upstream, packaging curated, stable knowledge agents need before runtime. It’s the difference between the socket and the content flowing through it.

The idea traces back to an April 2026 gist by Andrej Karpathy. He showed how Markdown libraries serve as “LLM wikis” that AI agents can read, maintain, and update themselves. Google Cloud formalized that pattern into a standard, ending years of fragmented, bespoke implementations named AGENTS.md, CLAUDE.md, and others.

OKF is out in the wild now. Version 0.1 is minimal and usable. It invites the ecosystem to build producers and consumers, expand the format, and shape how AI agents consume organizational knowledge. For any team serious about agentic AI, OKF is the standard to watch—or adopt—before your knowledge stays stuck in silos a while longer.