How to Connect ChatGPT to a Real Video Editor with MCP

What MCP actually is

The Model Context Protocol is a standard way for an assistant to call tools that live somewhere else. Instead of a model being limited to what its provider built in, you point it at a server, it discovers what that server can do, and those capabilities become available in ordinary conversation.

For video this is the whole ballgame. ChatGPT can write a script, art-direct a set of shots and tell you exactly where a cut belongs. What it cannot do is hold a project, render 1,800 frames, and hand you an MP4. Those are different problems, and MCP is how you attach the second one to the first.

What the connection can and cannot do

Worth being precise, because it is your account.

A connected assistant receives a token scoped to that one service. With it, it can read and change your projects and spend your credits on rendering and generation. It cannot read your email address or password, change your account credentials, or reach anyone else's data.

Everything else you discuss with the assistant stays between you and its provider. The server only ever sees the instructions the assistant chooses to send it.

You should not be pasting an API key

If a set-up guide tells you to generate an API key, copy a client ID out of a dashboard, or paste something from a database, that guide is describing an older and worse arrangement.

A properly built MCP server implements dynamic client registration (RFC 7591). The assistant registers itself automatically the first time it connects, which is why you have no client ID to enter. You paste one URL, you are sent to the service to sign in and approve, and you are done.

This matters beyond convenience. Under the old arrangement, only whoever had access to the underlying infrastructure could produce the credential — which meant everyone else simply could not connect. A URL that works for one person and not for anyone else is not a connector.

Connecting, step by step

1. Get the server URL. One URL, the same for every user.
2. Add it as a connector. In ChatGPT: Settings → Connectors → Add. In Claude: Settings → Connectors → Add custom connector. Paste the URL and leave the other fields alone.
3. Approve. You will be sent to the service to sign in and authorise. That issues the scoped token.
4. Verify. Ask it to list your projects. If it answers, you are connected.

You will need a plan that supports connectors — the option does not appear otherwise.

Directing once you are in

The point of attaching an editor rather than a generator is that the vocabulary gets specific. These are instructions, not prompts:

• Split that clip at 1.4 seconds
• Bring the headline in word by word, 40ms apart
• Duck the music under the voiceover
• Warm the grade half a stop and add a vignette
• Freeze the last frame for 800ms

Each of those is executed exactly or refused with a reason. Nothing lands halfway.

Revoking access

Any connection you can grant you should be able to remove in one click, without it affecting your other sessions. Check that the service lists its connections with a last-used timestamp — if you cannot see which assistants hold access and when they last used it, you cannot really manage it.

Treat an MCP connection the way you would treat any OAuth grant: connect the assistants you use, revoke the ones you do not, and check the list occasionally.