Specification v0.2 · Compiler v1.3.1 · experimental

Compiled Knowledge Format Documentation

The full reference for the file format, the v1.3.1 compiler pipeline, the MCP server and the integration surfaces that turn human knowledge into agent-ready context.

Start here

Getting started

Compile your first .ckf via the web UI, via MCP, or programmatically.

Schema reference

Metadata header + 22 typed sections that make up a CKF package.

Compiler pipeline v1.3.1

preflight → segment → chunk → reduce → promote → sanitize → coverage → numeric guards → canonical metadata → quality.

Preflight & coverage

v1.2 source preflight, record-level segmentation, coverage modes, numeric integrity, language recovery.

Language lock

Field-aware sanitizer + automatic language recovery.

MCP Server

One endpoint to expose CKF tools to Claude, Cursor, Windsurf, AI SDK, custom GPTs.

BYOK providers

5 providers (OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter) — keys never stored.

What is a CKF package?

A CKF package is a single file (.ckf.md, .ckf.yaml or .ckf.json) that describes a body of human knowledge as a metadata header plus 22 typed sections — entities, concepts, rules, procedures, retrieval chunks, atomic units and more — each carrying explicit confidence and source-basis labels so agents can reason over it without re-parsing prose.

Portable

Source-traceable

Record-traceable

Numeric-preserving

Agent-readable

Schema-stable

Language-locked

Where to compile

/compiler — full pipeline with chunking + reduce, web UI, BYOK or admin allowlist.
/compiler/demo — single-chunk LLM demo, anonymous-friendly.
/api/mcp — MCP Streamable HTTP endpoint for agents and IDEs.
/lab — A/B/A studies comparing CKF vs raw text on any document.
/news — every published article ships a .ckf companion + "Talk to this article".

Status

Specification: v0.2 (experimental). Protocol wire format: ckf-1.0. Reference compiler: v1.3.1. See Versioning for the full policy and the v0.1 → v1.3.1 timeline.