How to make a promo video from Claude or ChatGPT, and what to check before you approve it

The promoaivideos team · Updated

In short

Connect a video service to your assistant as an MCP connector, then ask for the video in chat. Claude and ChatGPT both ask before a connector takes an action, and that approval is your purchase order: check the format, length, script, voice and price before you allow it. Write the brief once, keep it in the chat, and reuse it for every product. A 15-second voiced video on promoaivideos costs 1,010 credits ($10.10) and takes 15 to 20 minutes.

Need a Promo Video?, mascot + product ui. Made with promoaivideos. See more videos

Your assistant already knows the product; the video lives elsewhere

If you write landing pages, changelogs and launch tweets with Claude or ChatGPT, your assistant already holds the raw material for a promo video: who the product is for, the one problem it solves, the feature that makes people say yes. What it cannot do on its own is turn that into a finished MP4. That job goes to a video service, and the bridge between the two is a connector.

This matters more in late 2026 than it did a year ago. OpenAI's API deprecations page lists the Videos API and every sora-2 model as shut down on 24 September 2026, with "no one-to-one replacement". If your plan was to wait for ChatGPT to make the video itself, that plan is gone. The working route today is the Model Context Protocol (MCP): a video service exposes tools, your assistant calls them, and the file comes back to the chat.

The thesis of this post is simple. Connecting the tool takes five minutes. The part that decides whether the video is any good, and whether you spend what you meant to spend, is the moment your assistant asks permission. Treat that prompt as a purchase order and you get a useful video on the first try.

How the connection works in plain words

An MCP server is a list of actions a service offers, such as "estimate a video", "make a video" or "edit a video". When you add the server to your assistant, it reads that list and decides when an action fits your request. You talk; it fills in the arguments and calls the tool.

The MCP specification is explicit about who stays in charge. Its tools chapter says there should always be a human in the loop with the ability to deny tool invocations, and that apps should show which tools are exposed, flag when a tool is called, and present confirmation prompts. Both Claude and ChatGPT do this, in slightly different ways, and the differences are worth knowing before you spend money through them.

Set it up in Claude, ChatGPT or Cursor

In Claude, Anthropic's help centre says custom connectors over remote MCP work on Free, Pro, Max, Ultra, Team and Enterprise plans, with Free accounts limited to one custom connector. On Pro and Max you add it under Customize, then Connectors, then Add custom connector, and paste the server URL. On Team and Enterprise an Owner or Primary Owner adds it in Organization settings first; members then click Connect. You can switch connectors on and off per conversation from the plus button in the chat.

In ChatGPT, OpenAI's developer documentation describes adding a custom MCP server from ChatGPT on the web, with a risk warning you must accept before it is created. It also says workspace permissions and security restrictions apply, so on a company workspace an admin may need to allow custom servers before you can add one.

For promoaivideos the steps are the same everywhere: open the setup page at www.promoaivideos.com/mcp, add the server to Claude, ChatGPT, Codex, Grok, Muse or Cursor, sign in once and click Allow. There is no API key to paste into the chat, and you pay with the same credits as the studio, with no extra fee. If you would rather wire it into your own scripts, there is a REST API on the same page.

The approval prompt is your purchase order

ChatGPT treats any tool that is not marked read-only as a write action and asks for confirmation by default, according to OpenAI's documentation. You can tell it to remember your decision for a tool within one conversation, and a new conversation asks again. Anthropic's help centre says Claude asks for approval before invoking tools, and warns you to choose Allow always only for servers and tools you trust to run unsupervised.

A paid video tool should never run unsupervised. The MCP specification adds that clients should show tool inputs to the user before calling the server. So read the inputs. Before you approve, check five things:

  • Format: 9:16 for TikTok, Reels and Shorts, 16:9 for YouTube and your site, 1:1 for feeds. One video per format, not one stretched to fit.
  • Length: the seconds you asked for. Lengths run from 5 to 30 seconds depending on the model, and longer costs more.
  • Script and on-screen words: your product name spelled right, your real price, no claim you cannot back up.
  • Voice: on or off. A voiceover adds cost; a sound-off feed video may not need one.
  • Price: promoaivideos shows the cost in credits before anything is made. If the number surprises you, deny and ask why.

Write the brief once and reuse it

Most disappointing AI videos come from a one-line request. Your assistant will fill the gaps with guesses, and guesses look generic. Keep a short brief in a project, a pinned note or a custom instruction, and start every video request by pointing at it. Here is a template that fits in one message:

  • Product: name, one sentence on what it does, the link.
  • Viewer: who is watching and where (a founder scrolling LinkedIn, a gamer on TikTok).
  • Hook: the first line on screen, written as the viewer's problem, not your feature.
  • Proof: the one screen, number or result you can show honestly.
  • Ask: the single action at the end (try it free, join the waitlist, shop the drop).
  • Look: a style name and two brand colours; say if you want a mascot or real product UI.
  • Specs: format, length in seconds, voice on or off, music on or off.
  • Never: words, claims or competitors that must not appear.

A worked example with real numbers

Say you sell a scheduling app and want one vertical video for Shorts and Reels this week. YouTube's help centre says a Short can run up to 3 minutes in a square or vertical aspect ratio, so a 15-second 9:16 film fits with room to spare. You paste the brief and ask: "Make a 15-second 9:16 promo for this, with a voiceover, in a clean motion-graphics look."

Your assistant shows the price in credits before anything is made. On promoaivideos, a 15-second voiced video made with Opus 5.5, the default model, costs 1,010 credits, which is $10.10 at one cent per credit. You approve. While it renders, you ask the assistant to draft the caption and three title options, since it already has the brief in context.

The video comes back with the price on screen reading $29 instead of $24. You do not remake it. You ask for the change in a sentence; a small text or colour edit costs 20 credits, and bigger edits start at 100 credits. Every version is kept, so you can go back. Your total for the week is 1,030 credits. On the Starter plan, $39 a month for 3,900 credits, that leaves room for two more videos of the same kind plus edits, and unused credits never expire, even if you cancel.

Plan for the wait

Video is slow compared with text. An Opus 5.5 video typically takes 15 to 20 minutes, and the other video models range from a few minutes to about 30. Chat assistants were built around replies that arrive in seconds, so how the wait looks depends on the app you use.

The protocol is catching up. The 2025-11-25 version of the MCP specification added tasks, an experimental feature built for long jobs: the server accepts the request, hands back a task ID, and the client checks the status until it is completed, failed or cancelled. Until every assistant handles that smoothly, the practical habit is simple. Start the render, keep working in the same chat, and ask your assistant to check on the video when you come back.

Five guardrails worth setting today

Connectors are powerful because they act, which is also why the platforms warn about them. Anthropic's help centre says malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions. OpenAI warns that an untrusted server may trick ChatGPT into using tools in unintended ways. None of this is a reason to avoid connectors; it is a reason to set them up with care.

  • Add servers only from the vendor's own site, and check the URL before you paste it.
  • Never click Allow always on a tool that spends money. Approve each video.
  • Turn off paid tools during deep research. Anthropic says Claude can call connector tools automatically during Research without asking again.
  • Switch the connector off in chats where you do not need it, so it is not called by accident.
  • Keep the brief as the source of truth. If the assistant drifts from it, correct the brief, not just the video.

Turn it into a weekly routine

Once the first video works, the payoff is repetition. Every new feature, price change or launch is a short request away: "Same brief, new hook: we now export to CSV. 9:16, 15 seconds, voice on." One brief, one approval, one file, and every version stays on your account.

Start small. Make one video this week for the channel where your buyers already are, post it, and note what you changed in the second version. That record is worth more than any template, including this one.

Sources

  1. Get started with custom connectors using remote MCP (Claude Help Center)
  2. Connect a custom MCP server to ChatGPT (OpenAI developer docs)
  3. Tools, Model Context Protocol specification 2025-11-25
  4. Tasks, Model Context Protocol specification 2025-11-25
  5. Deprecations (OpenAI API docs)
  6. Upload YouTube Shorts (YouTube Help)
  7. Make videos from your AI: MCP and API (promoaivideos)

Questions