Plans come out of registry files rather than a model, and the response says which parts of your goal the graph could not match. health_check reports the registry counts, so you can judge how much is actually behind a recommendation before you act on it.
A read-only advisory server over a registry of workflow components, the documented relations between them, published routes and playbooks. Describe a goal and it composes a route through the graph, scores its confidence, and says what the graph does not cover.
- 18 registered tools over the component registry
- plan_workflow — a goal in, a route with steps, connections, a safety note and build controls out
- list_graph_components and explain_component — what the building blocks are and what each one does
- list_known_routes and get_route — the published paths through the graph
- get_playbook — a published pattern with implementation guidance
- get_stack_recommendation and get_relevant_docs — technology choices for a deployment context, and the reference material behind them
- review_workflow_design — a second pass over a design you already have
- export_build_brief — a deterministic build and test contract for an agent to work from
- health_check — the registry counts, so you can see how much graph is behind an answer
- Coverage accounting that reports unmatched demand instead of pretending the graph covers everything
Node.js 20 or later and pnpm to run it over stdio, or point a client at the free hosted endpoint https://mcp.orchestratemcp.dev/mcp. It holds no secrets, makes no model calls inside its tools, and never writes.
One command — npx tsx src/server.ts