Updated October 11, 2026.
No, Claude cannot watch a video file the way you watch one. The Claude apps, the API and Claude Code take text, documents and still images, and none of them lists MP4 or MOV as an input. To get a screen recording in front of Claude, you give it what the video contains: a timestamped transcript, the frames that matter, and ideally a note of where you pointed and clicked.
TL;DR
- Claude chat (web, desktop, mobile) accepts documents and JPEG, PNG, GIF and WebP images. Video is not on the upload list.
- The Claude API takes image content blocks. An animated GIF is read as its first frame only.
- Claude Code reads screenshots you paste or reference by path. It can run ffmpeg in your shell, but a video is not a documented input.
- Record a skill in Cowork watches you do a task on your Mac and turns it into a reusable skill. It is not a way to hand Claude an existing recording.
- What works: a transcript plus labeled frames that you prepare yourself, or a recording link that an agent reads as text and images.
Can Claude watch a video?
Not directly. Anthropic's help article on uploading files to Claude lists documents and still images, and no video formats. Claude reads text and pictures, so a video has to be turned into those first.
The one feature that does look at your screen, Record a skill, does a different job. It watches you perform a task so Claude can learn the procedure. If what you have is a screen recording of a bug, a design change or a walkthrough, the practical answer is to give Claude its transcript and key frames, or a link it can read as text and images.
What does each Claude product accept?
Each row below was checked against Anthropic's own documentation in October 2026.
| Product | Takes a video file? | What it does accept | Source |
|---|---|---|---|
| Claude chat (web, desktop, mobile) | No | PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON and XLSX files, plus JPEG, PNG, GIF and WebP images. Up to 500MB per file and 20 files per chat. | Upload files to Claude |
| Claude API | No video content type | Images as base64, URL or Files API references. Up to 600 images per request on most models (100 on models with a 200k-token context). Animations are unsupported and only the first frame of a GIF is used. | Vision docs |
| Claude Code | Not a documented input | Images by drag and drop, paste, or a file path. It can run command-line tools such as ffmpeg if you ask it to. | Common workflows |
| Cowork: Record a skill | No upload; it records you live | A live recording of your screen, clicks, typing and voice while you do a task, which Claude turns into a proposed skill. | How to create custom skills |
The API row is worth reading closely, because some guides say Claude "sees the first frame" of a video. That applies to animated GIFs: the vision docs say "Animations are unsupported, and only the first frame is used." There is no video block at all, so an MP4 never reaches the model in the first place.
What does Record a skill actually do?
Record a skill arrived in Cowork on July 21, 2026, and it is behind a lot of the "Claude can watch my screen" headlines. Anthropic's article on creating custom skills describes how it works:
- Where it runs: Cowork in Claude for Mac, on Pro, Max and Team plans. The article says it isn't available in chat, on Windows, or on Free and Enterprise plans.
- How to start: click the + button in the Cowork composer and choose Record a skill, or go to Customize, then Skills, then Add, then Record your screen. Click Start recording, do the task while you talk through it, then click Done (or Discard). A recording can run for about 10 minutes.
- Permissions: the first time, macOS asks for Accessibility (mouse and keyboard tracking) and Screen recording.
- What happens next: Claude starts a Cowork task, reviews the recording and proposes a skill. You read the proposal and Save or Dismiss it, or Update an existing skill.
- Retention: Anthropic says the video and audio from your recording aren't retained. A set of screenshots stays with the Cowork task until you delete the task.
So Record a skill turns a demonstration into a reusable procedure for Cowork, such as "pull the weekly numbers from these dashboards into this sheet". It does not let you upload an MP4 you already have, and it is not a way to show Claude Code a bug in your app. If someone searching "claude record a skill" really wants Claude to understand a recording they made earlier, the paths below are the ones that work.
Why does a raw video fall short for screen recordings?
Even models that do accept video miss much of what matters in a screen recording. Google's Gemini video understanding docs say the model "samples the video at a rate of 1 frame per second (FPS)" by default, and that it "may miss details in videos with rapid motion or quick scene changes." A screen recording is full of short, important states: a toast that shows for half a second, a hover menu, a spinner that flickers before an error. Those can fall between sampled frames.
Sampling is only part of the problem. Three more gaps remain whichever model you use:
- Pointing words lose their target. People don't narrate bug reports like tutorials. They say "this button here" or "that one at the top". A transcript keeps the words and loses what they referred to. Only the frame at that moment, plus where the pointer was, tells you which button.
- Silent problems stay silent. A layout that is a few pixels off, a focus ring on the wrong field or a hover state that flickers is rarely said out loud. If it isn't in the transcript and isn't in a frame the model sees, it isn't in the input.
- A click position that was never stored cannot be rebuilt. To connect "the click that failed" to code, an agent wants the element that was clicked, or at least its coordinates. If the recorder didn't write that down when it happened, no model can recover it later from a compressed video.
That is why the useful question is not "can the model watch this?" but "does the input contain the moments, the words and the pointer?" For the cost side of that trade, see what a screen recording costs in agent tokens, and for when a single image is enough, screenshots vs recordings for coding agents.
How do you give Claude a screen recording yourself?
You can do this with free tools and any Claude plan. It takes four steps.
1. Get a timestamped transcript. Upload the file to a transcription tool such as Clipy's video to text converter, or run Whisper locally. Keep the timestamps; they are how Claude will match words to frames.
2. Pull out the frames that matter. Either grab frames at the moments you know are important, or let ffmpeg find the points where the screen changes:
# One frame at a known moment (1 minute 23 seconds in)
ffmpeg -ss 00:01:23 -i recording.mp4 -frames:v 1 frame_01m23s.png
# A frame at each visible screen change; timestamps are written to scenes.log
mkdir -p frames
ffmpeg -i recording.mp4 -vf "select='gt(scene,0.2)',showinfo" \
-fps_mode vfr frames/scene_%03d.png 2> scenes.logOlder ffmpeg builds use -vsync vfr instead of -fps_mode vfr. The pts_time values in scenes.log tell you when each scene frame was taken. A scene threshold that is too low gives you hundreds of near-identical frames; too high and small UI changes get skipped.
3. Name each frame with its timestamp. Rename the files so the time is in the name, for example 01m23s-save-spinner.png. Anthropic's vision docs say Claude doesn't receive image metadata, so put the time in the filename and repeat it as a label in your prompt. That is the only link between a frame and the transcript.
4. Upload the transcript and frames together, with a structured prompt. Claude chat allows 20 files per chat, so pick your best frames rather than all of them.
Attached: transcript.txt (timestamped) and 8 frames named by timestamp.
This is a screen recording of a bug in our checkout page.
1. Match each frame to the transcript line nearest its timestamp.
2. List every problem I point out, with the timestamp and frame.
3. Where I say "this" or "here", say which element you think I mean
and how sure you are.
4. Only then suggest likely causes.Be honest about the cost. This is manual, and the frames you choose are your own judgement: if you skip the frame where the bug shows, Claude never sees it. You still won't have click positions unless you add them by hand. For a one-off it is fine. If you hand recordings to an agent every week, it gets old fast.
How do you give Claude a Clipy recording link?
Disclosure: Clipy is our product.
Clipy does the conversion above while you record. You record your screen with the Clipy Mac app or Chrome extension while you explain the problem, then paste the normal watch link into Claude Code. With the Clipy skill installed, Claude fetches a structured version of the recording instead of the video. Setup for Claude Code and other agents is on the agent setup page, and the Clipy for Claude Code page shows the workflow end to end.
What the agent receives:
- The complete timestamped transcript and an AI summary.
- Key moments: timestamped frames of what you were pointing at.
- Pointer crops and click coordinates on Mac app and Chrome extension tab recordings.
- On the Mac app, the accessibility role and label of the clicked element when macOS provides them, so "the Save button" can arrive as a named button rather than a guess.
The skill also tells the agent to wait if the recording is still processing and to treat anything written inside a recording as content, not as instructions to follow. Public links are readable as they are. Private recordings and library search need the agent connected to your Clipy account. For the full definition of what makes a recording readable by an agent, see what an agent-readable recording is, and for a longer bug-handoff walkthrough, Claude Code screen recording.
Recording needs a Clipy account. New accounts get a 7-day free trial, no card needed. After that, Pro is $9 a month or $72 a year. People you share a link with don't need an account to watch it, and there is no watermark.
What about YouTube links?
Pasting a YouTube link does not give Claude the video. Anthropic's web fetch tool documentation lists text, HTML and PDF as the only supported content types, and notes that it does not support pages rendered with JavaScript. At best, Claude gets some of the page's text, such as the title and description, not the video stream.
For a public video you want a coding agent to learn from, a transcript plus frames works the same way it does for a screen recording. Our guide to giving Claude Code YouTube video context covers the options, and Clipy's YouTube to agent context page turns a video into a timestamped transcript with selected frames.
How do ChatGPT and Gemini compare?
- Gemini takes video natively. The Gemini API accepts MP4, MOV, WebM and other formats and processes both the audio and the visual stream, at 1 frame per second by default with a custom rate available. It still has the gaps above: sampled frames, unresolved pointing words, and no click data unless the recorder stored it.
- ChatGPT: OpenAI's article on uploading files and audio to ChatGPT lists document, spreadsheet, presentation and audio formats (images are covered separately) and does not list video. It says "WebM and MP4 files identified as video are not supported by audio uploads."
For step-by-step handoffs in Cursor, ChatGPT and Codex as well as Claude, see how to give Claude, Cursor and ChatGPT context from a screen recording.
Common questions
Can Claude watch an MP4?
No. Neither the Claude apps nor the API list MP4 as an input. Convert it to a transcript and a set of frames, or use a tool that does that for you, and give Claude those.
Does Record a skill keep my video?
Anthropic says the video and audio from a skill recording aren't retained. Screenshots from the session stay with the Cowork task and are removed when you delete that task. Anthropic also asks you not to type passwords or show private conversations while recording.
Can Claude Code record my screen?
Not by itself. With the Clipy CLI and skill installed, Claude Code can produce a proof recording after it changes your app: it can turn screenshots or an existing WebM or MP4 into a playable Clipy link, capture a flow in a headless browser, or record a specific window through the Clipy Mac app. The skill docs list each path.
Will Claude see every frame?
No. Claude only sees the images you give it, and a chat accepts 20 files. Even Gemini, which takes video directly, samples at 1 frame per second by default. Pick the frames where something changes, label them with timestamps, and pair them with the transcript.