Connect your AI agent (MCP)

Sushibelt is an MCP server. MCP, the Model Context Protocol, is how AI agents such as Claude Code, Cursor and Claude Desktop call outside tools. Connect one to Sushibelt and it can search your library and Discover, list models, price a workflow, start runs and read what they made — with a team API key you control.

The endpoint#

The server speaks MCP over Streamable HTTP at one URL. Every request carries a team API key as a bearer token, the same keys as the API.

https://app.sushibelt.ai/api/mcp

Authorization: Bearer YOUR_KEY

Create a key#

Create a key for the agent in the app, under Team settings → API keys, and name it after the agent. Give it only the scopes its tools need (see Tools): runs:read and workflows:read to look things up, and runs:create only if it should start runs. Set a monthly cap. Team settings → Connect an agent shows the snippets below with your endpoint filled in.

Replace YOUR_KEY in each snippet with the key. Keep it out of files you commit.

Claude Code#

Run this once in a terminal. Claude Code connects over HTTP, so nothing else needs installing.

claude mcp add --transport http sushibelt https://app.sushibelt.ai/api/mcp --header "Authorization: Bearer YOUR_KEY"

Then ask for something, such as "make a sprite of a mossy stone idol with Sushibelt". Run /mcp in Claude Code to see the connection and its tools.

Cursor#

Add this to ~/.cursor/mcp.json, or to .cursor/mcp.json in a project. A project's file is often committed: keep the key out of it, or keep the file out of the repository.

{
  "mcpServers": {
    "sushibelt": {
      "url": "https://app.sushibelt.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_KEY"
      }
    }
  }
}

Other clients, over stdio#

Some clients, such as Claude Desktop, start an MCP server as a program on your machine rather than connecting to a URL. For them, the @sushibelt/mcp package runs locally and forwards every message to the endpoint above. It has no tools of its own. It reads the key from SUSHIBELT_API_KEY, and the app from SUSHIBELT_URL (default https://app.sushibelt.ai).

{
  "mcpServers": {
    "sushibelt": {
      "command": "npx",
      "args": [
        "-y",
        "@sushibelt/mcp"
      ],
      "env": {
        "SUSHIBELT_API_KEY": "YOUR_KEY"
      }
    }
  }
}

SUSHIBELT_API_KEY=YOUR_KEY npx -y @sushibelt/mcp

Tools#

The agent gets these tools, the same ones the assistant in the app uses. A tool needs the scope shown; a tool the key's scopes do not allow is refused.

run_workflow and the make_* tools start a run and answer with its id and the credits it is expected to cost. They do not wait for it: the agent reads the run with get_run until it completes or fails.

Resources and prompts#

Resources are read-only views an agent can fetch by URI, or a person can attach in clients that support it.

Prompts#

• make_asset — A starting prompt for making an asset from a goal in words.

Logging and charging#

• Every call is logged. Each tool call is recorded against your team as a call from an MCP client, and listed in the app under Team settings → Provider usage.
• Runs are charged like the API. A run an agent starts costs the same credits as one started with Run a workflow. The free tools spend nothing.
• The key's limits apply. Its scopes, its workflow list and its monthly cap hold for every call, as they do for the API, and so do your team's other limits. See Cap your spend.
• Revoke to stop. Revoking the key in the app stops the agent at once.