What is CKP?

The Claw Kernel Protocol (CKP) is an open specification for describing, composing, and interoperating autonomous AI agents. It defines ten primitives that cover everything an agent needs: identity, inference providers, communication channels, tools, composed workflows, persistent memory, execution sandboxes, behavioral policies, multi-agent coordination, and observability.

Agents are described declaratively through a YAML manifest (claw.yaml). The protocol provides a formal ABNF grammar, a JSON-RPC 2.0 wire format, and a claw:// URI scheme for addressing agent components.

How CKP relates to MCP

CKP is complementary to MCP (Model Context Protocol), not a replacement for it.

MCP standardizes how LLM hosts discover and invoke tools, resources, and prompts from external servers. It solves the tool integration problem. CKP operates at a different layer: it standardizes how autonomous agents are assembled, secured, and orchestrated as first-class runtime entities.

In practice, any MCP server can be referenced as a CKP tool source. CKP’s Tool primitive is a strict superset of MCP’s tool definition, adding sandbox binding, policy binding, and lifecycle metadata on top of what MCP already provides.

CKP (Agent assembly, security, orchestration)
  |
  |-- extends -->
  |
MCP (Tool/resource discovery)
  |
  |-- uses -->
  |
JSON-RPC 2.0
  |
  |-- over -->
  |
stdio | HTTP/SSE | WebSocket | Message Queue

Three conformance levels

Not every agent needs every feature. CKP defines three conformance levels so implementations can range from minimal embedded binaries to enterprise swarm deployments:

Level	Name	Required Primitives	Target Use Case
L1	Core	Identity, Provider	Embedded devices, simple chatbots
L2	Standard	+ Channel, Tool, Sandbox, Policy	Personal assistants, team bots
L3	Full	All 9 core primitives (Telemetry optional at all levels)	Enterprise swarms, multi-agent systems

A valid L1 agent needs only an Identity and a Provider — it can run on a $10 board with a local Ollama instance. At L3, agents coordinate in multi-agent swarms with persistent memory, composed skills, and full policy enforcement.

The conformance harness ships with 31 test vectors (13 L1 + 10 L2 + 8 L3) so implementations can prove their level.

Learn more about each level in the Conformance Levels page.