MCP server for coding agents

mm-mcp is an MCP server on stdio. It lets a coding agent, such as Meta’s Muse Code or any other MCP client, author a control sequence the way our own agent does: read a spec, draft a block graph, and have our deterministic tools check it. Here the agent in the client is the muse.

The server makes no model calls and no network calls. Every verdict comes from the same code our authoring loop runs, so the same graph always gives the same content id and result. The protocol side is the official MCP Rust SDK, rmcp (pre-existing, pinned); the tools are ours.

Connect

The repository’s .mcp.json already lists the server as mechanical-muse. Build it once, then start it from the repository root:

cargo build -p mm-mcp
make mcp                 # cargo run --quiet --locked -p mm-mcp

Tools

Tool	What it does
list_specs	Each spec’s name, summary, typed inputs and outputs, scenario count
get_spec	The task text our agent receives, the authoring rules and the block-graph JSON Schema
palette	Authorable block classes from the pinned catalog, with ports and parameters
validate_graph	Fast structural findings, and interface problems against a spec
emit_cxf	The CXF text and engine content id
verify_graph	The full check: “passes X of Y scenarios”, the stage and code, each scenario’s result, and the repair message our agent would get
run_ledger, compare_runs	A recorded run’s Scope 1 and 2 ledger, and the difference between two runs of the same window

get_spec shows scenario names but not expected values, and no tool returns a reference graph, so the task stays a task. verify_graph’s feedback does name a failing expected value; a graph repaired against it has been repaired against the scoring vectors and should be reported that way.

Learn more

• docs/mcp.md: protocol revisions, error codes and an example session.
• docs/coding-agents.md: how Muse Code and Claude Code, which both helped build the repository, are set up in it.