MCP server

Connect any MCP client to https://api.fous.com/mcp (Streamable HTTP), with your API key as a bearer header. Or set the key as FOUS_API_KEY and tell your agent set up https://fous.com/SKILL.md: it connects the server and saves the skill.

The MCP server and API keys work on every plan (plans). On Free, an MCP client runs and reads your methods; the build tool answers 403 plan_limit, and builds start in Studio.

Connect your client

claude mcp add --scope user --transport http fous https://api.fous.com/mcp \
  --header "Authorization: Bearer ${FOUS_API_KEY:?Set FOUS_API_KEY to your Fous API key}"

codex mcp add fous --url https://api.fous.com/mcp --bearer-token-env-var FOUS_API_KEY

{
  "mcpServers": {
    "fous": {
      "url": "https://api.fous.com/mcp",
      "headers": {
        "Authorization": "Bearer ${env:FOUS_API_KEY}"
      }
    }
  }
}

{
  "inputs": [
    {
      "type": "promptString",
      "id": "fous-api-key",
      "description": "Fous API key",
      "password": true
    }
  ],
  "servers": {
    "fous": {
      "type": "http",
      "url": "https://api.fous.com/mcp",
      "headers": {
        "Authorization": "Bearer ${input:fous-api-key}"
      }
    }
  }
}

npx -y add-mcp https://api.fous.com/mcp -g -n fous \
  --header "Authorization: Bearer ${FOUS_API_KEY:?Set FOUS_API_KEY to your Fous API key}"

A Claude Code project

Commit this .mcp.json. Each person's own FOUS_API_KEY fills the header.

{
  "mcpServers": {
    "fous": {
      "type": "http",
      "url": "https://api.fous.com/mcp",
      "headers": { "Authorization": "Bearer ${FOUS_API_KEY}" }
    }
  }
}

Your own agent, server side

An agent that calls tools with no person approving them should use ?read_only=true: no run_action, build or stop, and the other tools only read (connections lists and reads connections; workflows lists and reads workflows and their runs, without their webhook addresses).

Anthropic Messages API, with anthropic-beta: mcp-client-2025-11-20:

{
  "mcp_servers": [
    {
      "type": "url",
      "url": "https://api.fous.com/mcp?read_only=true",
      "name": "fous",
      "authorization_token": "<FOUS_API_KEY>"
    }
  ],
  "tools": [{ "type": "mcp_toolset", "mcp_server_name": "fous" }]
}

OpenAI Responses API:

{
  "tools": [
    {
      "type": "mcp",
      "server_label": "fous",
      "server_url": "https://api.fous.com/mcp?read_only=true",
      "authorization": "<FOUS_API_KEY>",
      "require_approval": "never"
    }
  ]
}

The tools

Each tool does what its endpoint does: methods as in Find and manage methods, run and run_action as in Run a method, build as in Build a method. methods, builds, workflows and connections take an action that picks the operation.

structuredContent is what the API returns; a run's data matches the method's output_schema. The text next to it is a one-line summary and JSON shortened to fit 24,000 characters; in a run's text, a list of two or more records is written as {columns, rows} (each field name once, null for a field an item lacks) to save tokens.

Slow work

After about 50 seconds, run and run_action return status: "running" and an id. Call get_run with it, or stop, instead of sending it again. build returns at once (send wait, up to 60 seconds, to hold the answer); follow it with get_run. get_run and stop take a run (run_…), a build (bld_…) or a workflow run (wfr_…). get_run answers a run that failed with status: "failed" and its error, and stop with the run or build as it now stands (a method run in progress is free; a workflow run's finished steps stay charged).

Workflows

workflows runs the workflows API from one tool, with the same checks, through action:

A run is a workflow run with a wfr_ id. It answers with its steps once it has ended; while it is still going it answers status: "queued", "running" or "waiting": call get_run with its id to wait for it, or stop to cancel it. On the read-only server, workflows has only list, get and runs.

Actions and sign-ins

• run only reads. A write method goes through run_action, marked destructive so the host can ask you first. So are build, stop, workflows, methods and connections, so the host can ask before it changes something.
• On connection_required, give the person connect_url. Or call connections with action: "connect" and site, and relay each step to the person until the connection is active (Connect an account). connections also lists and reads connections, names one or makes it its site's default (update) and disconnects one (delete).

Errors

A failed tool returns isError: true and the HTTP error object in structuredContent.error. MCP-only codes are in Errors.