An MCP server publishes the checkable competence to use its complex tools, so any agent can learn, validate, plan, execute, and trace them through one interface, with the result on the record.
Seven layers, five tools, one handshake, with the agent at the center, deciding what to call, and when. The numbers are the usual journey; the agent chooses the order.
MCP finds a tool. Pilotage makes the whole program checkable.
A SQL engine, an automation platform, a workflow processor: the pattern repeats.
A JSON Schema can say a field is a string. It cannot teach the grammar the string must be written in, the names it may reference, or the rules that make it meaningful.
Real programs are drafted, checked, repaired, and only then run. A bare tool call gives an agent one confident shot and an opaque failure.
Today each agent bakes in private knowledge of each system. Publishing that competence on the server (checkable, versioned, discoverable) makes it work once, for every agent.
Every payload below is real, captured from Harborview, a public example server whose language no model has seen. Click a stage, or step through with the arrows.
Complete one real job in Harborview, chapter by chapter (discover, learn, validate, run, deploy, trace) until the whole loop is yours. Every request and response is real.
Open the tutorial → ReferenceTwelve short concept pages for the ideas, and the full four-part specification (Parts II and III normative), schemas, and 54 conformance vectors for the exact rules.
Browse the reference → ProofThree models from three vendors take the same job in a language none had seen. Watch them converge, and watch a confident “Done” get caught by a fact-checker.
See the evidence →Pilotage doesn't make agents identical, and it doesn't promise any one of them succeeds. It makes the server checkable and the loop honest, so different agents, taking different paths, converge on the same verified result.
Read the proof →