How to edit video with Claude using MCP

What MCP actually is (in one paragraph)

MCP is a standard way for an application to expose its capabilities to an AI agent. The app publishes a list of tools — each with a name, a description, and typed parameters — and the agent decides which to call and with what arguments. It is not screen-scraping or clicking buttons in a browser: the agent calls the same functions the app's own interface calls.

For a video editor, that means the agent can list your projects, read your timeline, add and trim clips, generate captions from real transcription, apply effects, preview a frame, and start a render. It sees the project as structured data rather than pixels, so it can reason about it: which clip is longest, where the silences are, whether a caption overlaps a face.

Connecting Claude to Vidmoat

Vidmoat exposes its editor over a hosted MCP endpoint, so there is nothing to install and no local server to keep running. You authenticate once with an API key from your account and Claude can drive your projects from then on.

1. 1In Vidmoat, open Account → API keys and create a key.
2. 2Add the Vidmoat MCP server to your Claude client using the config below, pasting that key.
3. 3Start a new conversation and ask Claude to list your Vidmoat projects — if it comes back with them, the connection works.

{
  "mcpServers": {
    "vidmoat": {
      "url": "https://api.vidmoat.com/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_VIDMOAT_API_KEY"
      }
    }
  }
}

What to ask for, and what it does under the hood

The useful mental model: you are not talking to a chatbot that gives advice about editing — you are talking to something that performs the edit. Vague prompts produce vague edits, so the wins come from asking for outcomes rather than nudges.

A few requests that map cleanly onto real operations:

1. 1"Cut every um and uh out of this interview" — transcribes the clip, finds filler words with their timings, and removes those exact ranges, closing the gaps.
2. 2"Add karaoke captions in our brand font" — runs transcription for word-level timings, then places and styles the caption track.
3. 3"Reframe this for TikTok and keep the speaker centred" — changes the project to 9:16 and tracks the subject so they stay in frame rather than being cropped out.
4. 4"Duck the music under the voiceover" — applies real sidechain-style ducking rather than a flat volume cut.
5. 5"Render it at 1080p and give me the link" — starts the export and polls until the file is ready.

The part most guides skip: checking the work

An agent editing a timeline it cannot see will make confident mistakes. Text lands off-screen, captions overlap a face, a font is too small to read on a phone. This is the single biggest difference between a demo and a workflow you would actually ship from.

Two habits fix most of it. First, ask for a frame: Vidmoat can return a rendered preview image at any timestamp, so the agent can look at its own work before rendering — and so can you. Second, read the lint. Every edit returns warnings for overlapping text, off-canvas elements and unreadably small type. Ask the agent to fix every warning before exporting and the output quality changes sharply.

The honest limitation: taste is still yours. An agent will hold a shot too long, or cut on the wrong beat, and it will not feel the difference. It is very good at the mechanical work that makes editing slow — transcribing, finding filler, syncing, reframing, restyling captions across 40 clips — and it is not a substitute for deciding what the video should be.

When this beats editing by hand

Repetitive, well-specified work at volume. Cutting one interview by hand is fine. Cutting twelve, each needing the same filler-word pass, the same caption style, the same vertical reframe, is where describing the job once and having it applied to all of them wins outright.

Work that is mechanical but fiddly is the other case: syncing multicam angles by audio, ducking music under dialogue, or restyling every caption in a finished project because the brand colour changed. These are not creatively interesting and they are exactly what an agent does not mind doing.

Editing that is mostly judgement — a narrative cut, a comedic beat, a montage that has to build — is still faster by hand. Use the agent to clear the mechanical work out of the way, then make the creative decisions yourself.