Skip to main content
graph mcp serve runs graph as an MCP server over stdio. Any MCP client can then call your plans as tools.

​
Add graph to a client

Point the client at graph mcp serve, and pass --dir with the project that holds your .graph/ directory. Clients start servers in a directory of their own choosing, so without --dir graph may not find your plans.

​
What the client sees

  • One tool per plan, named plan_<identifier> and titled with the plan’s name. Its input schema is the plan’s input_schema, and its description is the plan’s description with the exemplars added. A plan whose requires_servers aren’t configured isn’t offered.
  • What a plan tool returns: the plan’s result when it succeeds, or {error, kind?, step?, cause?} when it fails. See finish.
  • Tools for writing and testing plans, listed in MCP server tools. When the client changes a plan with them, graph tells the client to reload its tool list, so a new plan can be called right away.

​
During a run

  • Progress. When the client asks for progress, graph reports each step as it starts and the tokens spent so far.
  • Cancellation. When the client cancels the call, graph stops the plan before its next step.
  • Questions. An ask step becomes an elicitation request, so the client’s user answers it. If the client doesn’t support elicitation, or nobody answers within 10 minutes, the step’s whenUnanswered setting applies.
  • Long runs. A plan can take longer than a client’s default tool-call timeout. Raise the timeout in the client if long plans are cut off.