Record with your AI agent

CaptureFlow's local MCP server lets an AI agent (Claude Code, Codex, Cursor, or any MCP client) record videos on your own computer: the web app you're running on localhost, your whole screen, or one app's window. It can add title slides and a voiceover, then hands you a link you can share with anyone.

Everything runs on your machine: the browser, the screen capture and the voice.

Set it up

You need Node.js 20 or newer.

Claude Code

claude mcp add captureflow -- npx -y @captureflow/mcp

Codex

codex mcp add captureflow -- npx -y @captureflow/mcp

Cursor, Windsurf and other MCP clients: add a stdio server to the client's MCP config:

{
  "mcpServers": {
    "captureflow": { "command": "npx", "args": ["-y", "@captureflow/mcp"] }
  }
}

The first time the agent records, a browser tab opens so you can sign in to CaptureFlow. To sign in ahead of time, run npx @captureflow/mcp login.

Ask for a recording

Talk to your agent the way you'd ask a teammate:

• "Record a demo of the new signup flow on localhost:3000, with a voiceover."
• "Run the test suite and record the terminal while it runs."
• "Record a walkthrough of the settings page, then write a short report on what changed."

What it can record

What	How the agent records it
A web app, including localhost	Runs a hidden browser in the background and drives it: clicks, typing, scrolling, navigating. You keep using your computer. Only the actions are recorded, not the agent's thinking time.
Your screen	Records a whole display as it is, while the agent works with its own tools.
One app's window	Records just that window, like Terminal, VS Code or Figma, on a dark backdrop.

The agent can cut to a full-screen slide (a title, a list of steps, a closing line) between scenes. Every slide and page becomes a chapter viewers can jump to.

Reports

After a recording, the agent can write a report on it: a short summary above the video of what it did, what it checked and anything that failed, plus chapters. You can edit both on the recording's page.

Voiceover

Narration is free on every plan and runs on your computer.

• Default: the Kokoro voice. The first narration downloads it once (about 500 MB on disk, a minute or two); after that a sentence takes a couple of seconds. Ask the agent to say it'll narrate when it starts, and the download runs while it sets up.
• Your own voice server: point CAPTUREFLOW_TTS_URL at any server with an OpenAI-style /audio/speech endpoint, such as Kokoro-FastAPI running locally or OpenAI itself (set CAPTUREFLOW_TTS_KEY). CAPTUREFLOW_TTS_MODEL and CAPTUREFLOW_TTS_VOICE pick the model and voice.
• The system voice: set CAPTUREFLOW_TTS=system to use macOS speech, Windows speech, or espeak-ng on Linux.

Pass settings to the server in your client's config, for example claude mcp add captureflow -e CAPTUREFLOW_TTS=system -- npx -y @captureflow/mcp.

Permissions by platform

macOS: browser recordings need nothing extra. The first screen or window recording asks you to allow Screen & System Audio Recording for the app running your agent (your terminal or editor). Turn it on in System Settings → Privacy & Security, restart that app, and ask again. While it's allowed, that app can record your screen, so only grant it to apps you trust.

Windows: no prompt. Screen and window recordings start right away.

Linux: screen and window recordings need an X11 session (on Ubuntu, pick "Ubuntu on Xorg" at the login screen) and xrandr; install wmctrl to record a single window. Wayland sessions can still record the browser.

A screen recording films everything visible on that display, including notifications. Close anything private first.

Save files instead of uploading

Add --save-dir <folder> to save each recording as an MP4 in that folder, with its report as a Markdown file beside it. Nothing is uploaded and you don't need to sign in.

Options

Option / variable	What it does
--show-browser	Show the browser window while recording (hidden by default).
--save-dir <folder>	Save MP4 files instead of uploading.
CAPTUREFLOW_VIEWPORT	The browser's size, 1920x1080 by default.
CAPTUREFLOW_FFMPEG	Path to ffmpeg. Otherwise the one on your PATH, or a bundled build.

Limits

Recordings stop at 30 minutes; browser recordings also stop after 5 minutes without an action. Uploaded recordings count toward your plan like any other recording. They don't use your plan's agent recording minutes, which are for the cloud browser.

To disconnect a computer, run npx @captureflow/mcp logout and revoke its token on the Devices page.