Prepublish MCP

What prepublish/prepublish-mcp does

The prepublish/prepublish-mcp MCP server reviews a YouTube script before it becomes a video. It accepts text supplied by the MCP client and returns analysis about the opening hook, overall structure, pacing, passages that may lose attention, authenticity signals, applicable YouTube policy areas, and estimated speaking time.

It does not retrieve videos, transcripts, or channels. Its attention results describe relative risk within the written draft; they are not predictions of retention, views, watch time, or revenue. The service is free and anonymous by default, with no account or API key needed.

How it works

The hosted service uses Streamable HTTP and is stateless. Its primary endpoint is https://mcp.prepublish.ai; https://prepublish.ai/mcp and https://mcp.prepublish.ai/mcp are aliases for the same handler. The server supports several MCP protocol revisions and negotiates a supported revision during initialization.

The prepublish/prepublish-mcp MCP server exposes six tools. AI-backed requests are processed remotely, while script_runtime performs local arithmetic. The runtime calculator uses speaking rates of 160, 181, and 201 words per minute, based on measurements from 349 videos.

Setup and configuration

For clients that support remote HTTP MCP servers, configure the server URL directly. Claude Code can add it with its HTTP transport command, while clients using an mcpServers object can use an entry with type set to http and the hosted URL. Claude Desktop uses a custom connector, and Codex CLI accepts the URL in its MCP server configuration.

Clients limited to launching local commands can use the repository's stdio bridge through npx. The bridge requires Node.js 20 or newer and forwards requests to the hosted endpoint. PREPUBLISH_MCP_URL can replace the bridge's upstream URL. A PREPUBLISH_TOKEN bearer token can raise usage limits, but it does not add tools.

Tools and capabilities

• audit_script analyzes a completed draft, returning hook, structure, and pacing scores, an attention-risk map, a rewrite, and an analysis ID. The full result is released by email, so this tool requests an email address.
• get_audit retrieves an earlier result using its analysis_id.
• audit_hook examines sentence-level attention pull, the opening's curiosity gap, payoff distance, a grade, and alternative rewrites.
• check_authenticity identifies reused, templated, or mass-produced content risks and reports the quotations that triggered each signal.
• policy_preflight identifies YouTube policy families relevant to the script and links each finding to YouTube's published guidance.
• script_runtime converts word counts to runtime ranges or runtime targets to word budgets without consuming AI-backed quota.

Limitations and notes

The default limit is 100 AI-backed calls per UTC day across the endpoint and six AI-backed calls per caller per ten minutes. Runtime calculations do not count toward either limit. When the daily allowance is exhausted, the service returns an error with next-step guidance instead of asking the client to retry.

All checks are text-only and operate on the script passed to the tool. The service cannot assess the performance of a finished recording or provide a forecast of audience or business outcomes. The repository contains the stdio bridge and conformance tests; the hosted server implementation is part of the Prepublish web application.