Workflows as MCP tools¶

Superstep is an MCP server too. Turn it on for a workflow and MCP clients such as Claude Desktop, Cursor, Claude Code or other agents can call the workflow as a tool. This is the other direction from MCP servers, where your workflows call tools of servers you add.

Turn it on¶

Open the workflow, then More > API.
On Tokens, create a token. Clients authenticate with it, exactly as with the public API.
On MCP, switch on Expose as an MCP tool. It is off for every workflow until you turn it on, and only people who may manage the workflow's tokens can change it.
Pick your client and copy the config. Right after you create a token, the snippet already contains it; otherwise replace <token>.

The server is at https://<your studio>/mcp (Streamable HTTP, JSON responses, no session). Send the token as Authorization: Bearer wft_... or X-Workflow-Token: wft_....

.cursor/mcp.json

{
  "mcpServers": {
    "superstep-refunds": {
      "url": "https://superstep.example.com/mcp",
      "headers": { "Authorization": "Bearer wft_..." }
    }
  }
}

Claude Desktop starts local servers only, so its config runs mcp-remote to reach the HTTP server; Claude Code takes the URL directly:

claude mcp add --transport http superstep-refunds https://superstep.example.com/mcp \
  --header "Authorization: Bearer wft_..."

Tools¶

A token belongs to one workflow, so a client sees three tools:

Tool	What it does
`run_<workflow name>`	Runs the latest saved version. Its input schema is the workflow's state fields; the arguments are the run's input. Returns the run result: `status`, `final_state`, `error`, `waiting`, `metrics`.
`get_run`	Status and result of a run (`run_id`).
`resume_run`	Answers the Human node a waiting run is paused at (`run_id`, `response`: text, or `true` / `false` to approve / reject) and waits for the run to settle again.

Runs are ordinary API runs: they show in the workflow's history with source API, count against the token's rate limit (STUDIO_TRIGGER_RATE_LIMIT) and use the same input checks. A tool call waits up to STUDIO_SYNC_WAIT_MAX seconds; a run still going then returns status: running and its run_id, so the client polls get_run.

Human nodes¶

A run that reaches a Human node returns status: waiting with its run_id and the node's prompt under waiting. The client (or the person using it) answers with resume_run; the run continues from its checkpoint. People can also answer it in Superstep, in the editor or the inbox, as with any waiting run.

Errors¶

Failures come back as tool errors (isError: true) whose content is the common error object, for example {"error": {"code": "RUN_NOT_WAITING", ...}}. Requests without a valid token are refused before MCP with 401 TRIGGER_TOKEN_INVALID; a workflow that is not turned on answers 403 MCP_SERVER_DISABLED.

Demo mode¶

In the public demo the visitor account cannot create tokens or change settings, so it cannot expose a workflow.