Chat AIs like Claude, Codex and Cursor can reason about an edit, but they can't watch a clip on their own. This guide covers how to give an AI still frames from a tool on your computer, and how to use the notes without handing over the final call.
Disclosure: written by the WaveXML team. The tool used in the examples is ours.
What MCP actually does here
MCP (Model Context Protocol) is an open standard that lets an AI app call outside tools. The AI app is the chat you type in. The tool, called an MCP server, does a job and sends back an answer. The AI decides when to call it, then uses the answer for its next step.
Servers come in two kinds. A local server runs on your machine and answers only there. An online server runs on a company's computers, and you usually sign in to use it. For footage, that difference matters: a local server can read your media where it sits and hand the AI only the pieces it asks for.
The examples here use WaveXML, a free Mac and Windows beta with a local MCP server built in. It has no AI model of its own. It syncs cameras and recorders by sound, measures shots, makes still frames on request and writes edit files. Your AI makes the choices. The habits below also work with other MCP footage tools.
Connect your own AI
Add the server in your AI app, start a new chat, and make a test request. In WaveXML you can do it in either of two ways:
- Inside the app. Click Talk to your AI and your own Claude, ChatGPT (Codex) or Cursor opens in a chat in WaveXML. You sign in once with your own account, in your browser.
- From an app you already use. Click Connect… on the start screen and pick a row: Add to Claude, Add to Codex or Add to Cursor. For Claude Code, copy the one command shown and run it once in Terminal. In Cursor, switch WaveXML on under Settings, then MCP. Other MCP apps can use the details under Other apps.
The connection only works while WaveXML is open on that computer, and the AI only sees the tool in a chat started after you added it. The button at the top of WaveXML turns green and says AI while your AI is using it — a quick check that the connection works.
What an AI can and can't judge from stills
A still frame is a sample, not the shot. Stills are useful for framing (headroom, a cut-off chin, a mic in the shot), for who and what is in frame, for someone blocking the lens, for a shot that's clearly too dark, and for obvious softness if the tool sampled the right moment.
They are weak or blind for motion: shake, a whip pan or a slow drift are hard to see in a handful of frames. That's why it helps when the tool also measures shake and fast swings and reports the number. Stills also miss anything between the frames they sampled. And a sharp shot can still be the wrong shot — that call stays with you.
WaveXML pairs the two. It measures every shot for shake, soft focus, darkness and fast swings, with a number for each. It also makes frame sheets of any clip on request. Each frame is numbered and shows the camera and the time, so the AI can check faces and framing. WaveXML itself doesn't recognise people or faces; your AI looks at the pictures.
Prompting a shot check
Vague prompts get vague notes. Name the scope, the problems and the output you want:
Using WaveXML, check every shot from camera B during the speeches. List anything shaky, soft, too dark or a fast swing, with the time of each. Then show me frames of the three worst and tell me what you see.
Say "every shot", or you may get a sample. Ask for times, not adjectives: "soft from 00:14:02" is something you can check. Ask why a shot was flagged, so you can tell a measured number from a guess made from frames. If you already have an edit, ask for a marker timeline: WaveXML can write a copy with a note on each problem shot, which you import as a map and fix in your real timeline.
Review the notes like an assistant's
Treat the list as a report, not a verdict. The check can miss a problem, or flag a shot that's fine. Watch every flagged shot yourself, in motion, then skim the shots it didn't flag, especially the key moments. When it's wrong the same way twice, correct it and ask it to save a rule, for example: "From now on, always cut before a whip pan." WaveXML keeps these in My AI rules, and every AI you connect follows them. Only then let it cut a first edit — a draft to watch, fix and export to Resolve, Final Cut or Premiere.
Where your frames go
"Local" doesn't mean nothing leaves the machine. WaveXML's site says the MCP connection runs inside the app on your computer, your video files are never uploaded, and your AI only sees the still frames it asks for. WaveXML does not send your media or those frames to us. The frames and facts your AI reads (clip names, times, measurements) are sent by your AI app to its maker, for example Anthropic, OpenAI or Cursor, under that company's privacy policy. If a shoot is under NDA, check that policy before you ask for frames of sensitive moments.
Connect once, ask for times and frames, and watch everything it flags. The AI does the scanning. You still decide what goes in the cut.
Top comments (0)