Browser Use turns AI web tasks into polished demo videos

Browser-harness, an open-source browser control tool under browser-use, just merged PR #532 to give browser agents built-in screen recording and automated video editing. In short, while your AI clicks through websites and fills out forms, it records every move and automatically packages the session into a polished tutorial video that looks like it came out of Screen Studio.

What it does

Usually, tracking a browser agent leaves you digging through terminal logs or looking at static screenshots. This new update separates capture and editing into a clean two-step pipeline run entirely by the agent.

While running web tasks, the system takes per-action screenshots and saves an events.jsonl interaction trace. Once the task finishes, the agent invokes an internal make-video skill to turn those raw assets into a dynamic video.

The final output is far from a standard screen capture. It wraps the action inside a realistic Chrome browser frame with paper-textured gradient backgrounds, calm camera zooms, synthetic cursor motion, and crisp telemetry pills showing every step. If an action fails, the compositor adds a subtle red vignette and camera shake so viewers instantly see what went wrong.

It also takes privacy seriously. The recorder automatically strips OAuth codes, tokens, and API keys from URL parameters before writing to the event log. You can also define specific page coordinates to pixelate passwords or sensitive fields, and the blur tracks smoothly across camera zooms.

Under the hood and getting started

Instead of relying on heavy command-line tools like ffmpeg, the entire rendering pipeline runs directly in the browser.

The project uses a self-contained HTML compositor that draws frames, shadows, cursor sweeps, and titles onto an HTML5 canvas. It then pipes that canvas straight into the browser native MediaRecorder to spit out a WebM file. This keeps the codebase lightweight and removes the headache of system-level video encoding dependencies.

Recording is off by default so it does not clutter your disk. Telling your agent to record, show, or demo maps automatically to start_recording(). If you want to log everything, setting BH_RECORD=1 auto-records sessions and rolls them over after 180 seconds of inactivity to avoid blending distinct tasks.

To export a video, call the make-video skill or use the video CLI command included in the repository. Everything is already merged into the browser-use/browser-harness repo on GitHub, ready to run locally.

Why it matters

If you build browser agents, create software demos, or maintain internal product walkthroughs, this is a must-try. It takes a workflow that usually eats up half an hour across screen recording tools and video editors, and turns it into a zero-click byproduct of the agent running its task.

By leaning on the browser's own canvas and MediaRecorder instead of chunky encoding backends, browser-harness nails the visual polish without dragging in heavy infrastructure or sending your data to third-party rendering clouds. It gives browser agents a clear, human-friendly way to show their work.