MCP server

The Model Context Protocol is how an agent host is handed a set of tools, so it is how this kit intends to reach agent hosts.

Planned. This surface is scoped and not yet built. The decision that scoped it is .thoughts/decisions/2026-08-03-agent-facing-surfaces.md; decisions ship in the repository, so you can read the reasoning before any of this exists. There is no server to install and no command to run today.

What it will be#

A transport wrapper, and nothing more. The server contains no protocol logic and defines no tools of its own — it exposes the shared tool definitions with a name, a description and a JSON Schema input schema, as the Model Context Protocol specifies.

Stdio is the supported transport for local hosts. Any remote transport needs authentication and separate operational acceptance; hosted multi-tenant MCP is explicitly out of scope.

Failures split the way the protocol expects: an execution failure returns isError: true with actionable text, so a model can correct itself, while a protocol failure is a JSON-RPC error. Version and help must succeed with no wallet, no network and no key configured. Configuration — network, RPC, key — comes from the environment, and a secret is never written to disk or to logs.

Why the surface stays small#

An MCP server injects every tool schema into the host's context at startup, which costs materially more than a skill does. So the exposed surface is deliberately composed around user intents rather than one tool per protocol call.

The documentation knowledge surface ships separately from the transaction tool surface, so a host can adopt what the kit knows without adopting what it can do.

What exists today#

The headless operation contract the server would expose. It is already the layer the React hooks and widgets sit on — see the quickstart, and mock mode for driving it with no chain and no key.