Agentic video editing
Edit video by asking.
Tell the Assistant what you want changed. It turns the request into a plan, previews the edits on your timeline marked as AI changes, and you keep them, discard them or refine them.
In the browser today. Desktop apps for macOS and Windows are next.
The image behind this headline is generated, not footage. Nothing on this page is a mock-up of the editor.
We're bringing WebMCP to Elah, for ChatGPT and Claude.
The editor's Assistant already works through a single op layer: it describes the project, plans edits, and applies them as one undoable step. WebMCP will expose that layer, so an assistant you already use can drive the open editor directly instead of going through ours.
Planned. No release date.
What WebMCP is
- A proposal from the W3C Web Machine Learning Community Group, from Google and Microsoft.
- A page registers tools with document.modelContext.registerTool(), and an assistant in the browser can call them.
- Chrome shipped it in Canary 146 and ran an origin trial from Chrome 149.
- ChatGPT’s desktop browser and Codex discover these tools; Claude Code reaches them through an MCP bridge.
Designed to run on an i5 with integrated graphics.
You should not need a workstation to edit video. The editor is built to lean on what an ordinary laptop already has.
The browser’s own video decoder
Decoding goes through WebCodecs, so it uses the hardware decoder the browser already has. There is no WebAssembly codec running on your CPU.
Cheap work for an integrated GPU
Frames are composited with WebGL2 as textured quads: layers drawn as flat images, not a heavy 3D scene.
Export happens on your machine
Rendering runs in a Web Worker beside the editor, so the interface stays responsive, and nothing is uploaded or sent to a render server.
About 64 KiB gzipped
That is the open-source engine and editor UI as shipped to the browser, gzipped and minified. The video codec layer loads separately the first time you decode a file.
We don't publish benchmark numbers. These are the architectural reasons it stays light.
Assisted rotoscoping, with you correcting the frames that matter.
A first-pass mask for a shot, built around the way a roto artist already works: select on one frame, fix it by hand, and let the tool carry it to the next keyframe.
- 01
Upload one continuous shot
One shot at a time, checked in the browser before it is sent.
- 02
Box the subject on one frame
Draw a box around what you want cut out, on whichever frame shows it best.
- 03
Press Auto adjust
A first selection appears on that frame for you to judge.
- 04
Fix the edges with a brush
Paint in what is missing and erase what is not part of the subject.
- 05
Accept it as a keyframe
Your corrected frame becomes the reference the rest of the span follows.
- 06
Propagate across the span
The mask carries forward to the next keyframe, or to the end of the shot.
- 07
Correct a drifting frame
Fix it and accept it, and only its span runs again, not the whole shot.
- 08
Export the mask sequence
A PNG sequence, one mask per frame, ready for Nuke or After Effects.
In pilot with a VFX studio. Early access is through Request access.
Your footage is uploaded with media analysis opted out, so it is never queued for it, and it is deleted on a schedule.
The engine is open source.
The timeline, renderer and editor underneath Elah are published under the Apache-2.0 licence, as five packages you can install and build on.
- @elah/core
- @elah/react
- @elah/timeline
- @elah/editor
- @elah/cli
Edit your next video by asking.
Ask for early access to the editor, or write to us about the rotoscoping pilot. A person reads every request.