CHP and MCP: discovery, invocation, and evidence

If you're building agents, you've adopted or evaluated the Model Context Protocol. So the reasonable question about anything adjacent is: does this compete with MCP, or sit on top of it? For the Capability Host Protocol the answer is clear, and worth being precise about — they're complementary, and CHP should not claim to replace MCP.

Different questions

MCP and CHP answer different questions:

• MCP answers "what can the model call?" It's a stateful connect-and-invoke protocol — a client connects to a server, negotiates capabilities, lists tools (tools/list), and calls them (tools/call). It's model-facing, and it's becoming the standard way LLM applications reach tools and context.
• CHP answers "what actually happened, who was denied, and can I replay it?" It defines a capability boundary where every attempt emits structured, mandatory evidence — and it's agnostic about how the call was made.

One describes the surface a model talks to. The other records and governs what crossed the boundary. You can want both, because you usually do.

Three layers, not two competitors

It's cleanest to see the agentic stack as three layers, each answering its own question:

Discovery is a static, crawlable advertisement. Invocation is the live connection that runs the tool. Evidence is the durable, governed record of the attempt. MCP owns the middle layer well; CHP owns the third — and capabilities.txt hands off to whichever invocation layer you use.

How they compose in practice

The mapping is direct, not forced:

• An MCP server can be wrapped as a CHP host — its tools become declared capabilities that emit evidence on every call.
• A CHP capability can be exposed through an MCP-compatible tool surface, so a model can discover and call it normally.
• An MCP tools/call becomes a CHP invocation envelope carrying correlation, subject, and mode; the tool result becomes a CHP outcome — explicit success, failure, or denied — with evidence references attached.

The gaps CHP fills are exactly the ones MCP doesn't set out to cover: MCP doesn't require every tool attempt to emit started/completed/failed/denied evidence; its request IDs correlate a request to its response, where CHP correlation reconstructs causal execution across tools, agents, and replay; and MCP has no notion of a denial decision made before execution. None of that is a knock on MCP — it's a different job.

The honest summary

MCP exposes tools to AI applications. CHP governs and evidences execution of capabilities. Use MCP to let your agent reach its tools; wrap those calls with CHP when you need to prove what they did. The full concept-by-concept mapping — server↔host, tool↔capability, tools/call↔invocation envelope — is in the CHP vs MCP comparison.

If you're running agents on MCP today and the question "can we show what they did?" is coming, that's the layer CHP adds.