How to Connect Cursor to a YouTube MCP Server

Cursor adds Prepublish in one entry, either by hand or in one click. The by-hand route is a file, and there are two of them: .cursor/mcp.json for one project and ~/.cursor/mcp.json for every project. Paste this into either one:

{
  "mcpServers": {
    "prepublish": {
      "url": "https://mcp.prepublish.ai"
    }
  }
}

Save the file and restart Cursor. Prepublish audits a script you have written but not recorded, and Cursor is the client that puts that check next to the file you are writing the script in. The shared reference, including the six tools and the published limits, is the Claude Code hub guide. This page covers only what is specific to Cursor.

Cursor documents both install routes in Model Context Protocol and in MCP integrations, read on 2026-09-21.

One click. Open Customize in the sidebar, click MCPs, find the server and click Add to Cursor, then follow any authentication prompts. Prepublish needs no authentication, so nothing follows the click.

By hand. Create mcp.json in one of the two documented locations:

• .cursor/mcp.json inside the project, which the help page suggests committing so teammates inherit the same tools.
• ~/.cursor/mcp.json in your home directory, for every project on the machine.

The help page states one rule that matters more than the rest: both files are merged, and when the same server name appears in both, the project-level entry wins. That precedence is the cause of most confusing Cursor MCP behaviour, and it is the first thing to check when a server looks misconfigured.

The MCP reference adds what a URL entry can carry. A remote server uses url, and headers is where a token goes. Interpolation works in command, args, env, url and headers, with ${env:NAME} for an environment variable, so a key never has to be written into the file. Use it with a pp_live_ key:

{
  "mcpServers": {
    "prepublish": {
      "url": "https://mcp.prepublish.ai",
      "headers": { "Authorization": "Bearer ${env:PREPUBLISH_API_KEY}" }
    }
  }
}

Cursor also supports a local stdio entry, which is the shape the bridge takes. The documented fields are type, command, args, and env or envFile, and envFile is available for stdio servers only.

You may have seen a cursor://anysphere.cursor-deeplink/mcp/install link offered by a server's own documentation, usually with a base64 encoded config in a query parameter. That scheme was not verified on a Cursor-owned page on 2026-09-21. The two pages read for this guide, the MCP reference and the MCP help page, describe one-click installation as browsing Customize > MCPs and clicking Add to Cursor, and describe custom servers as entries in mcp.json. Neither documents a cursor:// install scheme.

Treat the deeplink as unverified in both directions: it is not documented on that page, and no claim is made here that it fails. You do not need it either way, because the file entry above is documented and the deeplink would write it.

Cursor supports all three transports: stdio, SSE and Streamable HTTP, per the MCP reference, so both routes are available to you.

Use the hosted URL by default. It is one line, it needs no Node installation, and it is the entry that survives a reinstall of the editor.

Use the bridge when something upstream forces the command shape:

• A Team or Enterprise allowlist may approve stdio servers by command pattern and remote servers by URL entry pattern, per the MCP reference. If your administrators approve commands rather than URLs, npx -y prepublish-mcp is the entry that will pass.
• You want the tool list served from a process you started, with the upstream named in its stderr line on startup.

The stdio entry is the same one every client uses:

{
  "mcpServers": {
    "prepublish": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "prepublish-mcp"]
    }
  }
}

The bridge forwards to the same hosted server, and prepublish-mcp/test/bridge.test.ts asserts that its sorted tool names equal the six names in the hosted list. It needs Node 20 or newer, per the engines field in prepublish-mcp/package.json. PREPUBLISH_MCP_URL points it at another deployment and PREPUBLISH_TOKEN forwards a bearer token, so the environment block is where a pp_live_ key belongs if you take this route.

Four failures are Cursor's rather than the server's.

A stale project entry overrides the global one. Because both files merge and the project entry wins for a duplicate name, a prepublish entry left in one repository keeps overriding the global entry there. If the server behaves differently in one project, read .cursor/mcp.json in it first.

The editor has to be restarted. The help page states that you save the file and restart Cursor, and that a server relying on shell environment variables needs those variables visible to Cursor, which also means a restart after a shell profile change.

A tool that has not been approved never runs. Cursor asks for approval before using MCP tools by default, and the MCP reference describes the run modes: the default Auto-review mode lets allowlisted MCP tools run immediately and routes everything else through a classifier.

The logs are in the editor. The help page puts them in the Output panel, opened with Cmd+Shift+U, with MCP Logs selected from the dropdown. A JSON syntax error and a failed connection both surface there.

When the problem is on the server side, the response says so, and these strings come from the endpoint's own code.

The last four are tool-level errors, so they appear in the conversation and the model reads them, which is why the message tells the model what to do next rather than only what went wrong.

Open the Agent panel in the repository where you keep your drafts and ask for the free arithmetic first:

Use the prepublish server to tell me how many words I need for a twelve minute video, and how long a 2,400 word script would run.

The call shows up as a tool invocation with expandable arguments and a response. The payload carries speaking_rates_words_per_minute with fast, median and slow, a source string, a caveat string, word_count, runtime_minutes and word_budget, the last two each with the same three keys. The rates are the measured ones in prepublish-fe/lib/seo/wpm-data.ts. Computed from them, a 2,400 word script is 15 minutes at 160 words per minute, 13.3 at 181 and 11.9 at 201, and a twelve minute target needs 1,920 words at the slow rate, 2,172 at the median and 2,412 at the fast rate. Nothing is billed for that call, so it doubles as a connection test.

Then hand over an opening, which is the check Cursor is best placed for, because the draft is in front of you:

Audit this hook with the prepublish server and name the sentence that loses attention first: "Most creators lose half their viewers in the first fifteen seconds, and the reason is rarely the editing."

audit_hook takes hook_text, an optional niche and an optional email, and returns a hook object with evaluation_id, hook_text, niche, overall_score, grade, one_sentence_verdict, sentences, top_issues, rewrites, model_used and created_at. Each entry in sentences has its own attention read, so the flagged line is a real line from the draft rather than a summary of it. This one is AI-backed, and Cursor asks for approval before running it.

The payload repeats its own limit, and the model relays it: "This is a text-only check of an unrecorded script. It maps relative attention risk inside the draft. It does not measure or predict published YouTube retention, and it cannot account for delivery, editing, thumbnail, topic or distribution."

The check reads text. It does not watch a video, fetch a transcript or read your channel, and no result is a forecast of audience behaviour.

The free tier is published in prepublish-fe/lib/mcp/limits.ts, and the values in force for the current day are in the server card at https://mcp.prepublish.ai.

A full audit is released by email, so audit_script asks for an address and the backend rejects a request without one. The narrower checks do not ask. An optional pp_live_ API key, interpolated from an environment variable as shown above, takes the call out of the shared daily cap and applies your account's plan quota instead.