Prepublish speaks MCP. Install it as a ChatGPT plugin, or connect it to Claude or any MCP client, and your assistant can check the script you just wrote (hook, structure, pacing, the passages most likely to lose viewers, and a rewrite for each) without leaving the conversation.
No account, no key, no card required. The free tier is a published daily allowance of AI-backed calls; the length calculator is local arithmetic and never counts against it.
In ChatGPT: install the plugin
Prepublish is a listed ChatGPT plugin, so there is nothing to configure. Open Plugins in the ChatGPT sidebar, search for Prepublish, and select Install plugin. No sign-in is asked for. It works on the web, desktop and mobile apps, on every ChatGPT plan that can install plugins.
Once installed, ask for an audit in plain words or mention @Prepublish. The full audit renders inline as a report card: scores, the relative attention-risk map, the flagged passages and the first rewrite. Install from the directory rather than asking the chat to find the plugin: in-conversation plugin search is unreliable for recently listed plugins. Step by step: Prepublish in ChatGPT.
Everywhere else: connect the endpoint
One URL, Streamable HTTP, stateless. Paste it wherever your client asks for a remote MCP server.https://prepublish.ai/mcp is the same server and keeps working, so anything already configured stays connected.
Endpoint
https://mcp.prepublish.ai
Claude and other remote clients
Add a custom connector and give it the endpoint above. Nothing else is required: an anonymous connection works immediately. Adding an API key is what lifts the shared daily cap, and it takes one line of config. ChatGPT Developer Mode accepts the same URL for an unlisted or test build.
Desktop clients that need stdio
Use the bridge. It forwards to the hosted server, so the tool list stays identical. Setting its PREPUBLISH_TOKEN environment variable to a Prepublish API key authenticates it the same way.
Optional. Anonymous calls work, but they carry no account, so they share the endpoint’s daily allowance with every other anonymous caller. Create a key at prepublish.ai/dashboard and send it as a bearer token on every call. The key is shown once, when it is created, and can be revoked from the same page afterwards.
Anonymous. Counted against the shared daily cap on AI-backed calls, and against a per-connection allowance of 6 calls per 10 minutes.
With a key. Attributed to your account, not charged against the shared daily cap, and allowed 30 calls per 10 minutes on that connection. Your account’s own plan quota is the limit that applies.
Those are the defaults and they are published live in the server card. Muse (muse.ai/platform) stores the key in its Secure Credentials Store and sends it on every call, so nothing has to be pasted into a prompt.
The six tools
audit_scriptAudit a finished draftCounts against the daily allowance
When
The user has written a script and wants to know what is weak before recording.
Returns
Hook, structure and pacing scores, a script-level attention-risk map naming the passages most likely to lose viewers, and a copy-paste rewrite for each flagged passage.
A previous audit was still running, or the user refers back to one from earlier in the conversation.
Returns
The current state of that audit, and a link to open it in the browser.
Inputs
analysis_id
audit_hookScore an openingCounts against the daily allowance
When
The user shares only a hook or first paragraph, or wants a fast read before auditing a whole draft.
Returns
Per-sentence attention pull, curiosity gap and payoff distance, an overall grade, the top issues, and rewritten alternatives.
Inputs
hook_text, optional niche, optional email
check_authenticityCheck reused-content riskCounts against the daily allowance
When
The user worries the script reads as mass-produced, templated or repetitive.
Returns
A score, a risk level, the firing signals with quotes, and remediation steps.
Inputs
video_title, script_text
policy_preflightPre-flight monetisation riskCounts against the daily allowance
When
The script touches sensitive subject matter and the user asks whether it is safe to monetise.
Returns
The YouTube policy categories the script may touch, each citing YouTube’s own published page, plus a verdict and flag counts. Quoted passages and their rewrites are part of the paid audit.
Inputs
script, optional title
script_runtimeLength to runtime, and backAlways free
When
“How long will this run?” or “How many words for a ten-minute video?”
Returns
A runtime range from measured speaking rates (160 / 181 / 201 words per minute across 349 videos), or the word budget for a target runtime.
Inputs
script_text, or word_count, or target_minutes
Try it without a client
The endpoint answers plain JSON-RPC over POST, and a bare GET returns a machine-readable server card.
It does not predict published retention. Every check reads text only. It maps relative attention risk inside an unrecorded draft; it cannot account for delivery, editing, thumbnail, topic or distribution, and no result is a forecast of what viewers will do.
It does not decide monetisation. The policy pre-flight reads your script against YouTube’s published guidelines and links the official page for each category. Only YouTube decides, and it decides on the finished video.
It does not write your script. These tools audit a draft you already wrote. If you want one generated, use a writing tool and bring the draft back here.
It does not invent an email address. An address is optional on a free full audit: give one and the result link is emailed to you as well, leave it out and the audit still runs. On a connection authenticated with an API key your account address is used.
Limits and data
Free tier
A published daily cap on AI-backed calls, resetting at 00:00 UTC, plus a short per-connection limit so one client cannot spend the day’s allowance. The live numbers are in the server card. When the allowance is gone the tool says so and tells your assistant not to retry.
Your script
Scripts are processed to produce the audit and stored with the analysis so it can be reopened. Nothing is sold or shared. See the privacy policy.