Can Claude Code read a video? 4 ways to give Claude Code a video (2026)

You recorded the bug. Thirty seconds, your voice explaining it, the cursor hovering over the thing that is wrong. Then you paste the file path into Claude Code and it tells you, politely, that it cannot watch videos.

That is still true in October 2026. Claude Code reads text and images. Native video input is an open feature request (anthropics/claude-code #12676, and a later one about screen recordings for UI bug reports, #80865). So every way to give Claude Code a video is really a way to turn the video into something it can read. The four below differ in what survives that translation.

What a coding agent actually needs from a recording

Before comparing tools, it helps to name the three things a screen recording carries that a text ticket does not:

1. What you said, in order. "This button should say Upgrade, not Submit."
2. What you pointed at while you said it. "This button" only means something if the agent knows which one.
3. When things happened. A flicker at 0:07, a total that changes after the second click. Order is evidence.

Any method that drops one of these makes your agent guess. Keep that list in mind for each option.

Way 1: paste screenshots

Pause the video at the frame that shows the problem, take a screenshot, and drop it into Claude Code. It reads images well, and for a static problem (a misaligned label, a wrong price) this is often enough.

Keeps: what the screen looked like at one moment. Loses: your voice, the order of events, and everything between frames. Three screenshots and a paragraph of explanation is a text ticket with pictures, which is the thing you were trying to avoid.

Way 2: paste the transcript

If your recorder transcribes (most do now), copy the transcript into the chat. The agent gets your words in order, which is a real improvement over a hand-written summary.

Keeps: what you said, often with timestamps. Loses: what you were pointing at. "This one is wrong" with no image is the classic failure: the agent has to guess which "this", and it guesses with total confidence.

Three spoken asks, each with its own mark. A transcript alone keeps the words and drops the circles.

Way 3: a local skill that extracts frames and transcript

Several open-source Claude Code skills now do the extraction for you. The pattern is the same: point the skill at a video file or URL, it runs ffmpeg to pull frames, runs a speech model to transcribe the audio, and hands Claude both, lined up by time. If you want everything to stay on your laptop, this is the right family to look at.

Keeps: words, frames and timing. Loses: intent. The skill picks frames by time or by scene change, not by what you meant to point at, so the agent still has to work out which element in a busy frame is the subject. You also own the setup: ffmpeg, a transcription model, and the disk space for both, on every machine that runs the agent.

Way 4: an MCP server that serves the recording as data

The fourth way moves the work off your machine. You record with a tool that stores the transcript, the frames and your on-screen marks, and your agent reads them over MCP. This is how Vidmatic does screen recording for AI coding agents: you record with the Chrome extension, talk, and draw on the page with the pen while you talk. Your agent calls the recording tools and gets:

• A timed transcript of what you said.
• Every mark you drew, a circle, an arrow, a cross or handwriting, with the second you started it and the second you confirmed it, so each mark carries the sentence you were saying.
• An image of the screen with each mark drawn on it, so the agent sees which element you circled instead of guessing from words.

0:04 the words, 0:05 the circle. The mark carries the sentence that was spoken while it was drawn.

Keeps: words, timing and intent. When your words and your drawing disagree, the drawing wins, because people misspeak while they point and the mark is usually the precise part. Costs: a hosted service has your recording (your code stays on your machine; the agent reads your repository locally).

Side by side

What Claude Code does with it next

Reading the video is only the first step. Once the agent has the recording, the useful question is what it does with it. With the Vidmatic MCP, every recording has a Give this to your agent panel with one line to paste:

/mcp__vidmatic__workstream_from_video <recording id>

Your agent reads the transcript and each mark, searches your repository, and writes one issue with one row per ask and the file and line where it lands. If it cannot find an anchor it says so rather than inventing one.

The circled Submit button lands on src/app.js:10, and one issue is filed.

From there it builds, reports each stage live, and can record a before and after video of the fix. The whole loop, filmed on a real change, is in Circle it. Say it. See it shipped.

Which way should you pick?

• A single static glitch: a screenshot is fine.
• You want everything local and you are happy to maintain ffmpeg and a speech model: a local skill.
• You record bugs often, the bug is in where you are pointing, or you want the agent to file the issue and prove the fix: an MCP server that serves marks.

To try the fourth way, connect once and sign in in the browser; there is no key to paste. The screen recording MCP server page has the command and the list of skills.