Skip to main content
Beta NDI includes a hosted Model Context Protocol endpoint. It lets an AI agent start document jobs and read their results without the agent constructing REST requests itself.

​
Connect from Cursor

Add the server to your project’s .cursor/mcp.json or your user-level MCP configuration: 
{
  "mcpServers": {
    "nace-ndi": {
      "url": "https://ndi-api.nace.ai/mcp",
      "headers": {
        "X-API-Key": "${env:NDI_API_KEY}"
      }
    }
  }
}
Set NDI_API_KEY in the environment that launches Cursor, then enable nace-ndi under Settings → Tools & MCP. Ask your Nace representative for an API key if you do not have one.
Treat the API key as a secret. Do not paste it directly into a committed mcp.json file.

​
Available tools

  • parse_document — start converting a document to markdown.
  • categorize_document — classify document sections or spreadsheet sheets.
  • ground_items — locate requested values and return source coordinates.
  • get_job — read job status, result, or error.
  • cancel_job — request cancellation of a pending or running job.
Job-starting tools always use NDI’s asynchronous endpoints and return a job_id. The agent can call get_job until the job is succeeded or failed. The same account limits and credit charges apply to REST and MCP calls.

​
File inputs

The hosted MCP server cannot read paths on your computer such as /Users/me/report.pdf. Pass one of the sources accepted by the API:
  • a publicly reachable or presigned https:// URL;
  • an allowed s3:// URL; or
  • an ndi://file/<uuid> handle returned by POST /upload.
To process a local file, upload it through the REST API first and give the returned handle to the agent.

​
Beta contract

The MCP endpoint is currently an API-key-authenticated beta. Tool names and inputs are intended to remain stable, but OAuth and support for additional MCP clients may be added after customer feedback. The REST API remains NDI’s canonical public contract.