Cursor MCP setup: project config, HTTP and a read test

Start with a scope decision in Cursor: should this connection be available in this project only, or across your projects? Choose the server and access method afterwards. This order keeps personal configuration from being mistaken for a team default. The first example uses tracevero’s public registry search, which reads registry data and needs no tracevero access key. Success with this endpoint does not establish access to your other work applications.

Separate project configuration from global configuration

Cursor documents .cursor/mcp.json inside the project and ~/.cursor/mcp.json for global configuration. Both use mcpServers. A local stdio server takes a command; a remote server takes a URL. The Cursor client reference and transport comparison help classify details you already have.

Read an existing file before adding anything. Do not replace the entire object if it already contains other servers. Pick an unambiguous name and record which file changed. If your team versions project configuration, keep personal secrets out of it. Knowing the configuration path does not approve every operation that the server may later offer. Keep a local copy of the earlier configuration to make the change reversible.

A complete HTTP example for the first test

Put the following content into your chosen mcp.json, or add its entry inside the existing mcpServers object. The URL is an MCP endpoint. A search result or a web page from the server directory cannot substitute for that endpoint.

{
  "mcpServers": {
    "tracevero": {
      "url": "https://api.tracevero.com/mcp"
    }
  }
}

This example starts no local process. Another provider may require authentication. Cursor documents OAuth for supported servers and variables such as ${env:NAME} in configuration values. Use the provider’s intended authorization method. Possessing a token says little about actual access unless the account, scope and requested operation also match. Record account context without copying credential values into a report.

Check the server, tool and result separately

1. Check the file and its scope first. Keep your other server entries unchanged during the experiment so you can attribute the result to this addition.

2. Open Cursor’s MCP management and inspect the connection. Read a concrete startup or HTTP failure before replacing credentials.

3. For tracevero, choose the read-only search operation with the term filesystem. Inspect the offered arguments before confirming the call.

4. Open one result’s evidence link and compare it against the browser registry search. Record the query and timestamp: the total number of registry entries can change later.

Why a connected server can still be unusable

Keep four observations separate: configuration was loaded, the connection succeeded, the tool was offered, and the call returned the expected result. State exactly which stage succeeded in a report. For the last stage, choose a small answer you can independently verify. A long plausible response is less useful for acceptance than one correct name with its corresponding evidence. Repeat that bounded operation after any configuration change.

For absent tools, use the focused troubleshooting guide. You can also inspect a problematic server through the MCP Inspector. If that works, compare endpoint, account, environment and arguments. Only after those agree can you meaningfully investigate a client-specific difference. The converter helps with supported format changes and explicitly reports its limits. Follow the format conversion walkthrough and environment variable guide for these handoffs.

Can I copy a VS Code configuration directly?

Check the source format: .vscode/mcp.json uses servers, while portable .mcp.json and Cursor use mcpServers. Matching root keys alone are insufficient; transport fields and variables must also fit the target client.

Does a project file work for every teammate automatically?

No. Each person needs the right runtime and, where applicable, their own authorization or organization approval. Share instructions instead of personal credentials.

Why does a local command work only in my terminal?

Compare executable path, working directory and environment. The ENOENT guide helps isolate a missing startup executable.

Must I approve each tool call?

Check current Cursor settings and organizational policy. This walkthrough does not assume automatic approval.

Vendor sources checked on 1 October 2026. Test workflows are editorial suggestions.

1. Cursor: MCP configuration, scopes and tools
   Show retrieval command

   curl -s https://cursor.com/docs/mcp

All posts