Custom stdio MCP server not available in Automation running on My Machine

[nrudz (Nrudden)]

Hi Cursor team,

I’m trying to use a custom MongoDB MCP server with Cursor Automations running on a My Machine worker, but the MCP server is not available inside the automation session.

Setup:

• Cursor Automation runs on a My Machine worker hosted on EC2.

• Worker is online and can run repo commands.

• New Relic MCP works in the same automation (HTTP based).

• MongoDB MCP is configured in Cursor UI as a custom MCP server.

• MongoDB MCP is stdio/command-based using npx.

• The EC2 machine has stable Atlas access; Mongo connectivity works from the machine.

Mongo MCP config:

{
  "mcpServers": {
    "mongodb": {
      "command": "npx",
      "args": [
        "-y",
        "mongodb-mcp-server@latest",
        "--readOnly"
      ],
      "env": {
        "MDB_MCP_CONNECTION_STRING": "redacted",
        "MDB_MCP_READ_ONLY": "true",
        "MDB_MCP_TELEMETRY": "disabled"
      }
    }
  }
}

Problem:

• In the automation, the agent can see/use New Relic MCP.

• The agent says no MongoDB MCP server/tools are available.

• Running MCP list commands on the My Machine worker also does not show the MongoDB MCP server.

• Repo-local or machine-local MCP config also does not seem to be picked up by the My Machine automation session.

Questions:

1. Are custom stdio MCP servers supported for Cursor Automations running on My Machines?

2. If yes, where exactly should the MCP server be configured: dashboard integration, automation config, repo .cursor/mcp.json, or on the My Machine host?

3. Are dashboard-configured MCP servers executed on Cursor’s cloud side or on the My Machine worker?

4. Do My Machine workers currently support launching command-based MCP servers like npx mongodb-mcp-server@latest?

5. Is there any known limitation where remote HTTP MCP servers work, but custom stdio MCP servers do not appear in My Machine automation sessions?

[Colin]

Hey @nrudz!

This should work! I was able to get the MongoDB MCP server you mentioned running in an Automation on a My Machines worker.

MCP servers have to be configured in the Automation itself, not in a .cursor/mcp.json file. A My Machines automation session only loads the MCP servers listed in the Automation.

You add it under Tools → Add Tool or MCP → MCP Server → + New Connection → MongoDB (the below screenshot shows what it looks like after it has been added). Click on Manage to configure.

Once it’s listed in Automation, the agent can pick it up and use it. Here it lists the databases on the cluster:

A few other notes:

• stdio MCP servers run on your My Machines worker. So a stdio server (like mongodb-mcp-server via npx) needs node/npx available and the database reachable from the worker machine, while an HTTP server needs to be reachable from Cursor’s backend.
• If you run cursor-agent worker start --verbose, you can see logs about which MCP servers are making it to the automation (look for the Loaded session MCP servers line when a run starts).

[nrudz (Nrudden)]

@Colin

Loaded session MCP servers

Thanks for the quick response.

I have everything set up like that, but the worker is reporting no available MCP servers (both chat thread & server logs using —verbose), and the automation reports the same in the chat thread.

I have verified:

• npx & node installed and accessible on the worker machine
• Mongo access via the connection string works from the worker machine

I have tried various configurations of the MCP environment variables (interpolating secrets and hardcoding credentials), and all have the same result

[Colin]

Shot in the dark that it could be Linux related? Seems like there are some issues running this MCP server in Linux as a non-root user.

1.11.0 crashes on spawn with ERR_MODULE_NOT_FOUND for @mcp-ui/server

You might try pinning the version to 1.10.0, or installing globally (npm install -g [email protected]) and changing the command to remove the npx -y!

It’s worth a shot at least, to rule out issues with the install!

[nrudz (Nrudden)]

Ok, embarrassing honesty. I realised my EC2 instance had limited disc space remaining when I tried and failed to globally install the mcp server. Turns out that was the issue all along.

I’ve increased disc space, and the automation is now successfully dynamically loading the MongoDB MCP server configured in the automation.

Wonder if there’s a possibility of improving verbose logging if something like this fails whilst attempting to dynamically load an MCP server to the worker machine

Thanks for the support.