How to Connect Claude Desktop to a YouTube MCP Server

Prepublish serves a YouTube script audit over the Model Context Protocol at https://mcp.prepublish.ai. Claude Desktop connects to it two ways: as a remote custom connector, which needs no file and no key, or through the local config file with the npx -y prepublish-mcp bridge. Start with the connector. It takes one paste and reaches the same six tools as every other client.

The server audits a script you have written but not recorded. The shared reference, including the six tools and the published limits, is the Claude Code hub guide. This page covers what is specific to Claude Desktop.

Open Claude Desktop, click Customize in the sidebar, then Connectors, then the + button and Add custom connector. Paste https://mcp.prepublish.ai and click Add. On a Team or Enterprise plan an Owner adds it once for the organization under Organization settings > Connectors > Add > Custom > Web, and each member then connects to it from Customize > Connectors. Both paths are described in Anthropic's article Get started with custom connectors using remote MCP, read on 2026-09-21.

Prepublish needs no OAuth, so leave the optional Advanced settings fields empty. The same article states that Free plans are limited to one custom connector, so if that slot is taken, remove the other connector or use the local file route instead. Enable it per conversation from the + button at the lower left of the chat, then Connectors.

One property of connectors decides whether this route works for you. Anthropic's article states that Claude connects to a custom connector from Anthropic's cloud infrastructure rather than from your machine, in every Claude client, so a server behind a firewall will not connect even when your own browser can reach it. https://mcp.prepublish.ai is public, so the hosted route works. The same article notes that local servers configured in claude_desktop_config.json use your local network instead, and are not available in Cowork or claude.ai.

Use the file when you want the server process on your own machine or the connector route is unavailable on your plan. The click path is in the Model Context Protocol documentation Connect to local MCP servers, read on 2026-09-21: open the Claude menu in the system menu bar, choose Settings, then the Developer tab, then Edit Config. That opens the file, or creates it when it does not exist.

The documented locations are ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows. Put this block inside it:

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

Then quit Claude Desktop completely and start it again. That page is explicit that the file is read at launch and that invalid JSON fails without a visible error, which is the most common reason a server never appears.

If it still does not appear, the same page names the logs: ~/Library/Logs/Claude on macOS and %APPDATA%\Claude\logs on Windows, with mcp.log for connection events and mcp-server-prepublish.log for this server's stderr. It also lists the steps: restart the app, check the file syntax, and make every path absolute.

Anthropic also ships desktop extensions, a packaged local route installed as an .mcpb file from Settings > Extensions, or from Advanced settings > Extension Developer > Install Extension (Getting Started with Local MCP Servers on Claude Desktop, read on 2026-09-21). Prepublish does not ship an .mcpb bundle, so your two routes are the connector and the JSON file above.

The hosted URL and the bridge are not two servers. https://mcp.prepublish.ai is a stateless Streamable HTTP endpoint that answers one JSON-RPC message per POST. The bridge, published on npm as prepublish-mcp, is a local process that connects to that same endpoint and re-serves the result over stdio, the transport a desktop app uses when it launches a subprocess. prepublish-mcp/test/bridge.test.ts asserts that the bridge's sorted tool names equal the six names in the hosted list, so the two cannot drift apart.

The choice for Claude Desktop follows from the two mechanisms above.

• The connector route takes a URL and nothing else. Use it unless the next two conditions apply.
• The config file route can only launch a command, so it needs the bridge: the npx -y prepublish-mcp entry above.
• Use the bridge when the audit has to leave from your own network, because a custom connector connects from Anthropic's cloud.

The bridge needs Node 20 or newer, per the engines field in prepublish-mcp/package.json. On startup it writes one line to stderr naming the upstream URL, the tool count and whether a token was supplied, which is the fastest way to confirm that it launched and reached the server. PREPUBLISH_MCP_URL points it at another deployment, and PREPUBLISH_TOKEN forwards an Authorization: Bearer header upstream, which is how a pp_live_ API key travels.

Claude Desktop lists six tools from tools/list: audit_script for a finished draft, get_audit to collect one by id, audit_hook for an opening, check_authenticity for reused-content signals, policy_preflight against YouTube policy categories, and script_runtime for words-to-runtime arithmetic. The hub guide and the MCP server documentation cover them in full.

When a call fails, the message is the diagnosis. The strings below come from the server's own code.

Two of those are transport-level, so no MCP call was made: the origin and protocol-version refusals. The rest are tool-level errors that arrive inside the conversation, so the model reads them and can tell you what to do next. A request that carries no Origin header is allowed, and the 403 fires only for a present, unlisted origin. A connector that was added but not enabled for the current conversation does not appear in the tool list at all, so check the + then Connectors toggles before you debug the endpoint.

Start with script_runtime, which consumes no allowance and proves the connection end to end. Type this with the connector enabled:

How many words do I need for a twelve minute video, and how long would a 2,400 word script run?

The tool returns a JSON object with these fields, all written by prepublish-fe/lib/mcp/tools.ts: 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 carrying the same three keys. The rates are the measured ones in prepublish-fe/lib/seo/wpm-data.ts. Run the arithmetic on them and a 2,400 word script is 15 minutes at 160 words per minute, 13.3 at 181 and 11.9 at 201, while twelve minutes needs 1,920 words at the slow rate, 2,172 at the median and 2,412 at the fast rate. The caveat field carries the limit in the payload itself: speaking time only, with pauses, B-roll, demonstrations and edits adding time.

Then try the check the audit is built around. Paste an opening and ask:

Score this opening and name the sentence that loses people 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 carries its own attention read, so the answer is anchored to your lines. This one counts against the allowance.

Both payloads repeat the same constraint, and Claude Desktop shows it because it is part of the response: "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 every number is a relative read on the draft rather than a forecast of what an audience will do.

All three counters come from prepublish-fe/lib/mcp/limits.ts, and the values in force 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, created at prepublish.ai/dashboard, takes the call out of the shared daily cap.