Skip to content

Knowledge Context Protocol (KCP)

The Knowledge Context Protocol (KCP) is an open standard that makes a body of knowledge navigable and trustworthy for AI agents. It is a YAML file format — passive data, not executable config — that describes the knowledge units in a project: their topology (depends_on, supersedes), their intent (what question each unit answers), their freshness (validated dates), their audience, and the trust evidence behind them. An agent can traverse a KCP manifest to find exactly what it needs without loading everything into its context window first.

If Model Context Protocol (MCP) is how an agent calls tools, KCP is how an agent finds and trusts knowledge. The two compose: KCP provides structure, MCP provides retrieval.

I have been designing KCP in the open since January 2026. This page is the map; the posts below are the territory.

At a glance

v0.24 · twenty-four versions in six months (v0.1 shipped January 10, 2026) · six layers — discovery, meaning, trust, time, identity, federation · reference agent shipping on npm · published under Cantara and submitted to the Agentic AI Foundation (Linux Foundation).

Why not just llms.txt?

llms.txt gives an agent a table of contents — a flat list of files. That is a real improvement over nothing, but a table of contents cannot express topology (what depends on what), freshness (is this still true?), intent (what question does this answer?), audience (who is this for?), or trust (who signed it, and can I verify that?). Agents need maps, not tables of contents.

KCP is that map. It started as a fix for a specific, embarrassing failure — a 33-tool-call bug where an agent kept rediscovering the same context — and grew, one falsifiable release at a time, into a full protocol for the agentic web.

Does it actually work?

KCP is an engineering claim, so it ships with numbers. Manifests measurably cut how much an agent has to read and how often it has to guess.

Evidence Result
KCP on Two Repos, Two Days 119 → 31 tool calls on application code; 53 → 25 on documentation. A 53–74% reduction in agent work.
KCP on Three Agent Frameworks Same pattern holds across Claude Code, CrewAI, and others — the gain is in the data, not the harness.
We Cancelled a 45-Minute Architecture Review "What else breaks if we change this API contract?" — a meeting, replaced by a 1.2-second query.

The protocol, layer by layer

Every layer has one job. Read the tour, or the release that built each one.

Why passive data matters: a manifest an agent reads cannot inject instructions the way config an agent executes can. That single design decision is what makes KCP defensible — see Why KCP Is Passive Data, Not Executable Config.

The tooling

KCP is a spec, but it ships with working implementations you can install today.

Governance & compliance

Because KCP carries trust and provenance, it turns regulation into something an agent can actually read.

Reading guide

KCP is developed in the open and still evolving. The blog has the latest releases; this page has the map.