Sunra

You run this MCP server to expose Sunra.ai models and tools to code assistants and editors. It provides a unified interface that lets tools list models, fetch schemas, submit jobs, and stream results, all while keeping your API key secure in environment variables.

How to use

Start the MCP server to enable editor and IDE integrations. Use the HTTP transport for remote access or run in stdio mode for local tooling. Connect your MCP-enabled clients (like Cursor or Claude Desktop) to the server to list models, fetch model schemas, submit jobs, and stream results.

For Cursor, point the MCP server URL to the server endpoint and provide your Sunra API key in the environment. Then use tools such as list-models and model-schema to explore and interact with Sunra models directly from your editor.

How to install

Prerequisites you need before starting: a Linux/macOS or Windows environment with Node.js installed and access to an internet connection.

Install and run the MCP server using NPX (no local build required):

npx @sunra/mcp-server --transport http --port 3925

Configure clients to connect to the MCP server

Set your Sunra API key in the environment before starting clients. Use the same key you use in other Sunra integrations.

For Cursor integration, use this MCP server config example to point Cursor to the local server:

{
  "mcpServers": {
    "sunra_mcp_server": {
      "url": "http://localhost:3925/sse"
    }
  }
}

Security and environment

Keep your API key secure. Set the API key in your environment and avoid exposing it in source control. The MCP server uses environment variables to manage credentials.

Examples and quick reference

Two common ways to run and connect are shown here. Use the HTTP method for remote clients and stdio mode for local tooling.

Available tools

list-models

Lists available Sunra models and tools exposed by the MCP server.

model-schema

Fetches the input/output schemas for a selected Sunra model.

subscribe

Submits a job to a Sunra model and streams results when available.

subscribe_async

Performs an asynchronous subscription to model results when supported.