Screen Recording for AI Coding Agents

Show your coding agent the screen instead of describing it. It reads what you said, what you circled and when, with your code open.

  1. 1Connect Claude Code (once per machine)
    claude mcp add --transport http --scope user vidmatic https://api.vidmatic.ai/mcp

    No key in it: run /mcp, choose Authenticate, and sign in from your browser.

  2. 2Record your screen, then hand your agent the recording
    /mcp__vidmatic__workstream_from_video <recording id>

    Use the id from your recording's link. Your agent reads it and files one GitHub issue.

Loading demo…
One real change, filmed end to end: a pricing page with three bugs, circled and talked through, read by a coding agent with the code open, built, and proven with before and after video in ten minutes.

Watch on its own page

How it works

1

Record and mark it up

Record your screen, circle what is wrong and talk through what you want.

2

Your agent reads it over MCP

Claude Code gets the timed transcript, every mark you drew with the second you drew it, and images of what you circled.

3

One issue, then the fix

It files one GitHub issue grounded in your code, works it, and you watch it live on Agent Workstreams.

4

Proof comes back

A narrated demo of the shipped change lands in your Inbox.

  • A two-minute recording is a better spec than three paragraphs
  • Your agent reads your marks and words, not a guess at a video file
  • Nothing to install beyond one connect line

How screen recording for AI coding agents works

Your coding agent cannot play a video file. It reads text and images. Vidmatic turns your recording into exactly that, serves it over MCP, and leaves the code to the agent that already has it.

What you record

To record your screen for Claude Code, Cursor or any other agent, use the Vidmatic Chrome extension when position matters. Talk while you record, and when you reach the problem pick up the pen: circle the wrong button, cross out what should go, or write the right value by hand. Each mark is saved as an image at the moment you confirm it.

No extension, or the problem is in a desktop app? Press New Recording in the Vidmatic app and pick any screen or window. There is no pen in that mode, so say out loud which element you mean and pause on it for a second. Either way, a thirty-second take with your voice replaces the paragraph of "the grey button on the middle card, after the second click".

What your agent receives

When you hand the recording over, your agent calls get_recording_intel and gets three things. The transcript, in short timed segments. Every mark you drew, with the moment you confirmed it and the sentence you were saying then. And frames: for each mark, get_annotation_image returns the screen with your mark on it, so the agent looks at the exact element you circled.

When your words and your drawing disagree, the drawing wins, because people misspeak while they point and a circle does not. If the transcript is still processing, the agent waits the number of seconds the response gives and asks again. Nothing about video runs on your machine: no ffmpeg, no downloads, no local model.

Which coding agents it works with

Vidmatic is a hosted MCP server at https://api.vidmatic.ai/mcp with sign-in in the browser. Claude Code and Cursor are the clients we verify end to end. In Claude Code you add it with the one line above, and every workflow is a slash command.

Screen recording for Cursor works the same way: add the vidmatic entry to ~/.cursor/mcp.json, start the sign-in from Cursor's MCP settings, then ask the agent to call get_skill with the workflow name, such as workstream_from_video. A video bug report for Cursor then reaches the agent as words, marks and images, just as it does in Claude Code.

Screen recording for Codex, or for any agent we have not listed, is not something we have verified. Vidmatic is a standard remote MCP server over HTTP, and its browser sign-in has to return to your own machine, which is how Claude Code and the Cursor desktop app do it. A client that works the same way may connect and reach the same tools through get_skill, but the MCP server page lists only the clients we have checked end to end.

The loop, in six steps

1. Circle it: record your app with the extension, draw on it and talk.

2. Every mark, timed: Vidmatic pins every word and every mark to the second, so each mark carries the sentence you were saying while you drew it.

3. Hand it off: paste the line from the recording's Give this to your agent panel. Your agent reads the recording, searches your repository and files one GitHub issue with a row per ask, each with the moment, the mark and one to three file and line anchors, or "anchor: none found" when it cannot find one.

4. Watch it work: your agent reports each stage, plan, build, review and ship, to Agent Workstreams, so you can see which issue it is on and where it is stuck.

5. Prove it: on request it records the bug before the fix and the same path after it against your running app, and Vidmatic pairs the two halves per requirement.

6. Ship it: the pull request merges, and the narrated proof lands in your Inbox next to the issue it answers. The film above is this loop on one real change, start to finish in ten minutes.

How it compares

Clipy, Cap and Jam also hand a recording to a coding agent, and each is good at something. Clipy turns a recording into a structured brief and captures console errors. Cap is open source and lets an agent run the recorder. Jam is built around console and network data. Vidmatic's difference is what you pointed at: every mark you draw reaches the agent as data with its time and an image, and the loop closes with a grounded issue, a live view of the work and before and after proof. The Vidmatic vs Clipy vs Cap comparison below lays it out row by row, with dates and sources.

Your code stays with your agent

Vidmatic holds the video; your agent holds the code. It reads your repository on your machine and sends back only what it writes on purpose: the issue text, stage summaries and the demo videos it records of your running app. Recordings stay private to your account until you share them, and your agent signs in as you to read them.

Questions

Can Claude Code watch a screen recording?
Not as a raw video file. Through the Vidmatic MCP server it reads your recording as a timed transcript, the marks you drew with the second each was drawn, and images of what you circled.
What is video-to-agent?
Video-to-agent is the loop Vidmatic runs: you record your screen, your coding agent reads it, files one GitHub issue, ships the change and sends back a narrated demo as proof.
How do I record my screen for Claude Code?
Connect Claude Code once with the claude mcp add line and a browser sign-in. Record with the Vidmatic Chrome extension, or press New Recording in the app, then paste the line from the recording's Give this to your agent panel into Claude Code.
Can I send a screen recording to Cursor?
Yes. Add the vidmatic entry to ~/.cursor/mcp.json, sign in from Cursor's MCP settings, and ask the agent to call get_skill with workstream_from_video. It reads the same transcript, marks and frames Claude Code does.
Which coding agents does it work with?
Claude Code and Cursor are verified end to end. We have not verified Codex or other agents. Vidmatic is a standard remote MCP server over HTTP with a browser sign-in that returns to your machine, so a client that supports that may work, but the MCP server page lists only what we have checked.
Do I need to install anything?
One claude mcp add line, once per machine. You sign in from your browser; there is no key to copy and no local ffmpeg.
Does my source code go to Vidmatic?
No. Your agent reads your repository on your machine and sends back only what it writes on purpose, such as the issue text and stage summaries. Recordings stay private to your account until you share them.
  • Screen Recording to GitHub Issue

    Turn a screen recording into a GitHub issue your agent can build: numbered requirements tied to the moment you said them, grounded in your code.

  • AI Agent Proof-of-Work Demos

    Proof of work for AI coding agents: when your agent says done, Claude Code records your running app and narrates the change, so you can see it before you merge.

  • Screen Recording for Claude Code

    Screen recording for Claude Code: connect Vidmatic with one line, record and circle the bug, and Claude Code reads your words, marks and frames over MCP.

Show Your Agent the Screen

Connect once, record once, and hand your agent the recording.

Explore more Vidmatic use cases

Docs: What your agent receives from a recording