macOS recorder · agent lane · one editor

Give your product an agent that films its own demos.

Describe the flow. It drives, records, and cuts the video. Or record your own screen and let the editor do the zooming.

Download for macOS

curl -fsSL https://www.clankcut.com/install.sh | bash

take 03 — onboarding flow — 00:41

mp4 / 60 fps

zoom 1.8×

camera keyframes

narration bound to steps

musicducked −14 db under vo

the editor — auto-zoom, wallpaper framing, mp4 out

60fps

screen capture

8ms

camera sampling grid

120

max agent steps

0.5s

between keyframes

ai demo mode

Anything you can demo, an agent can film.

Point it at your web app, give it one sentence, and it drives Chrome itself. The agent reads your page through its accessibility tree — it clicks the actual “Sign in” button, not coordinates.

  1. 01Open the URL you gave it in Chrome
  2. 02Read the page through its accessibility tree
  3. 03Click the actual “Sign in” button, not coordinates
  4. 04Type, wait, scroll, verify — up to 120 steps
  5. 05Record the run with a timestamped event track
  6. 06Drop the take into the same editor you already use

stock chrome · web apps · macOS app for the recorder

https://app.example.com/settings/team

role=button · name=“sign in”

accessibility tree

  • · document
  • · navigation
  • · main
  • · form “team invite”
  • · textbox “email”
  • button “Sign in”
  • · link “forgot password”

no pixel guessing

what the editor already knows

It edits like someone who watched you work.

01 / auto-zoom

Zoom that follows the click

The camera zooms to your clicks — not motion guessed from pixels. A physically-simulated spring drives every push-in, so it settles like an operator instead of snapping like a keyframe.

02 / cursor

Smoothed pointer, real ripples

On a macOS screen recording, the pointer path is smoothed from the timestamped input track and every click lands a ripple. The video and the event track are captured side by side.

03 / narration

Narration anchored to steps

The voiceover is written from the execution trace and attaches to steps, not timestamps. Cut a clip and the narration goes with it.

04 / presenter

A bubble and a score

Add an AI presenter bubble over the frame. Add AI background music. It ducks itself under the narration so nobody has to draw a volume envelope.

05 / takes

Every take is versioned

Re-run the flow, keep the old cut. Takes stack under the same video, so picking the one that landed is a click and not a file-naming exercise.

06 / durability

Jobs survive the app closing

Queued, leased, retried. Quit mid-run and the work is still in the queue when you come back — nothing lives only in a window that is still open.

how it works

Three steps. One MP4.

01

Record it, or describe it

Hit record in the macOS app and drive the product yourself, or hand the agent a URL and one sentence: “sign in, create a project, export a clip.”

02

It captures the run

60 fps video plus a timestamped input-event track. The agent lane drives Chrome through the accessibility tree and produces the same two artifacts.

03

The cut lands in the editor

Auto-zoom from real clicks, wallpaper framing, narration bound to steps, music ducked underneath. Export MP4 and send it.

agent-native

Your coding agent can queue the demo.

One install drops in the CLI and an agent skill, so the thing that shipped the feature can also file the demo for it. Jobs survive the app closing — queued, leased, retried.

Download for macOS

terminal

pay as you go

$ curl -fsSL https://www.clankcut.com/install.sh | bash
# clankcut cli + agent skill installed

$ clankcut recordings create \
    --url https://app.example.com \
    --flow “sign in, create a project, export a clip”

queued · leased · take lands in the editor

Questions, answered.

A Mac. Clank Cut is a macOS app, and the AI demo agent drives web apps in stock Google Chrome.

Point it at a URL and describe the flow in a sentence. The agent drives Chrome itself — reading the page's accessibility tree, not guessing at pixels — records the run, and the take lands in a project like any other recording. If it gets stuck, it stops and tells you instead of guessing.

Every recording carries a timestamped input-event track. Your clicks cluster into zoom regions, and a spring-simulated camera moves between them — the export is exactly what you previewed, frame for frame.

Yes. Trim, split, re-speed, adjust zoom regions, swap the wallpaper framing, regenerate narration. On recordings you make yourself you can also restyle the cursor — size, smoothing, click ripples.

AI narration is written from what actually happened and anchored to steps, so cutting a clip keeps the voiceover in sync. Background music generates to fit the timeline and ducks itself under speech.

Yes — one command installs the clankcut CLI, the agent skill, and on a Mac the desktop app itself if it's missing. Recordings queue as durable jobs that survive the desktop app being closed, with retries handled for you.

An MP4 — H.264 at 60 fps with a keyframe every half-second, so scrubbing is instant anywhere you drop it.

Pay as you go, billed per organization. No subscription tiers, no seats.

Don’t record the demo. Describe it.

Zero recording. Real demos. macOS app for the recorder, Chrome for the agent, one editor for both.

Download for macOS