OKF
OKF for Claude: Give Claude Your Real, Current Knowledge Graph over MCP
August 5, 2026 · 7 min read
Pasting documents into a chat is lossy, stale, and unverifiable. OKF gives Claude something better: a live, linked, cryptographically signed knowledge graph it can traverse over the Model Context Protocol — so its answers come from your real, current data.
Claude is only ever as good as the context you give it. Paste a document into a chat and you get a lossy snapshot: no links between facts, no freshness, no way to prove where an answer came from. Open Knowledge Format (OKF) fixes that — it lets you hand Claude a live, linked, verifiable knowledge graph instead of a wall of pasted text.
This post is the practical version: what OKF is, why it pairs so naturally with Claude, and how the two work together over the Model Context Protocol (MCP).
What OKF is, in one paragraph
OKF is an open, vendor-neutral standard for packaging knowledge as a graph of linked Markdown files with YAML frontmatter. Each file is a node — a control, an asset, a policy, a person — and the links between them are the edges. Because it is just Markdown, a human can read it, Git can version it, and — crucially — an AI model can traverse it natively. For the full primer, see what is Open Knowledge Format.
Why Claude and OKF fit together
Claude reads Markdown fluently and, over MCP, can call tools to fetch exactly the nodes it needs rather than swallowing one giant blob. That is the whole point of a graph: instead of stuffing everything into the context window and hoping, Claude follows links — from a question, to the relevant node, to its neighbours — and answers from the specific facts that matter.
- Grounded, not guessed — answers trace to real nodes in your graph, not the model’s memory.
- Current, not stale — the graph is regenerated as your data changes, so Claude never reasons from last quarter’s snapshot.
- Verifiable — an OKF bundle can be cryptographically signed, so you can prove what Claude was actually looking at.
- Portable — the same bundle works with any OKF-aware tool; you are not locked into one assistant.
“Pasting docs gives Claude a photocopy. OKF gives it the filing cabinet — indexed, cross-referenced, and stamped with the date.”
How it works, end to end
The flow is four steps, and only the last one involves Claude:
- Compile — your source of truth is rendered as linked Markdown: one node per entity, edges as links.
- Sign — the bundle is hashed and Ed25519-signed, so it is a verifiable, point-in-time record.
- Deliver — sync it to a Git repo or publish it where your tools can reach it.
- Consume — point Claude at the bundle over MCP and ask questions in plain English.
That last step is where the OKF-vs-MCP distinction matters: OKF is the *format* (how the knowledge is packaged) and MCP is the *protocol* (how Claude reaches it). They are complementary — OKF is what MCP carries.
A concrete example: compliance
CATAAM uses exactly this pattern. It compiles your live compliance program — controls, assets, automated tests and attack-surface findings — into a signed OKF bundle, and exposes it to Claude through the CATAAM Claude connector. Then you can ask:
“Which controls cover this S3 bucket, and are they passing right now?” Claude walks from the asset node, to the controls that reference it, to the latest test results — and answers from your real, current state. No dashboard-digging, no stale export.
One safety note for the plumbing: when you wire an assistant into real systems, keep secrets out of the model context. Prompt Guard handles that so credentials never ride along into the prompt.
OKF vs the alternatives for feeding Claude
| Approach | Fresh? | Linked? | Verifiable? |
|---|---|---|---|
| Paste documents into chat | No — manual snapshot | No | No |
| Vector RAG over docs | Only if re-indexed | Weakly (by similarity) | No |
| OKF bundle over MCP | Yes — regenerated on change | Yes — explicit graph edges | Yes — signed |
Vector RAG and OKF are not mutually exclusive — but where your knowledge has real structure (compliance, org data, product catalogues), an explicit graph beats similarity search, because the relationships *are* the answer. More on giving agents grounded context in OKF for AI agents.
See Claude answer from your real, current compliance graph — CATAAM compiles it to a signed OKF bundle and connects it over MCP.
Explore the OKF Context Engine →Frequently asked questions
- How do I use OKF with Claude?
- Produce your knowledge as an OKF bundle (linked Markdown with YAML frontmatter), then expose it to Claude over the Model Context Protocol (MCP) with a connector. Claude reads the Markdown graph natively and answers by traversing links to the relevant nodes, rather than you pasting documents into the chat. With CATAAM, you enable the OKF Context Engine and point the Claude connector at the bundle.
- What is the difference between OKF and MCP for Claude?
- OKF is the format — how knowledge is packaged as a linked-Markdown graph. MCP (Model Context Protocol) is the transport — how Claude reaches tools and data. They are complementary: MCP is how Claude connects, and OKF is the well-structured knowledge it connects to. Using them together gives Claude grounded, current, verifiable context.
- Why is OKF better than pasting documents into Claude?
- Pasting is a lossy, stale snapshot with no links between facts and no proof of provenance. An OKF bundle is a live graph: it is regenerated when your data changes, its nodes are explicitly linked so Claude can follow relationships, and it can be cryptographically signed so you can verify exactly what the model was looking at.
- Is OKF specific to Claude?
- No. OKF is an open, vendor-neutral standard, so an OKF bundle works with any OKF-aware AI tool or agent, not just Claude. Claude is a natural fit because it reads Markdown well and supports MCP, but you are never locked into a single assistant.