Cursor works because code is text.
A model can read the whole repo, change one line, and you can check the result. The models didn't get a special upgrade to make that happen. The material was already in a shape they could handle.
Video isn't. In most editors the timeline is a project file only that editor understands, and the footage under it is just pixels and audio. An AI can watch an export and tell you what it thinks. It can't reach in and move a clip.
That's the gap we built Arsaze around.
What Arsaze is
Arsaze is an AI video editor that you can use by hand and that an assistant can operate just as well. Cutting, trimming, moving clips and searching footage all work the same whether you click or an assistant asks. The editor in your browser is where you see it happen.
The same reach for you and the assistant
Most editing software was made for a person with a mouse. AI showed up later as a panel or a menu item. That works until the AI needs to do something the menu doesn't have.
We wanted the assistant to have the same reach as you. Adding a clip, cutting it, moving it, trimming it and finding a scene all work whether you do them by hand or ask an assistant to do them, so an assistant can do with the editor what you can do with it.
Claude, ChatGPT, Gemini and Grok can all connect over MCP and work on a real timeline today. Not a demo reel, the actual surface.
A timeline that knows what's in it
A list of clip names and timecodes isn't enough. What an editor reasons about is content: who's speaking, what's on screen, what just got cut.
Arsaze's timeline carries what is in the footage, not just clip names, and it updates as you edit. Trim a clip and what is now on screen and in the audio changes with it. Cut the sentence where the speaker names the client, and the AI knows the name is gone from the cut.
I'm keeping the how to ourselves for now. The what is simple: the AI works from what's really in the cut, not from what was in the raw file.
Finding the shot
An agent that can't find the right footage is guessing. So everything you add to Arsaze, video, audio and images, can be searched by scene or by spoken word.
Most of the value here is boring, which is fine. "Find every take where she mentions pricing" used to eat an afternoon. Now it's one query, whether a person types it or an agent does.
A request like "take this interview, pull every answer about pricing, cut it to a tight piece, caption it" maps straight onto things Arsaze already does: search, cut, arrange, caption.
Watching it happen
Open a project in the browser and you can watch an AI edit it live, including Claude connected over MCP. Cuts land. Assets get searched. The playhead jumps around.
We put a lot of effort into making this look good, probably more than was strictly reasonable. Waiting on an AI is easier when you can see what it's doing.
For developers
If you build with MCP, an assistant you connect can use the editing features you use by hand, including cuts, captions, color, audio and export. That's the part I'd want if I were building on top of a video tool.
The manual tools are still there
Nothing here takes the editor away from you. Rewrite lets you edit the video by editing its transcript. Auto Cut strips silence, or specific words you name. Auto Captions works down to individual words. Instant Voice Cloning takes under a minute, and Voice Over records in the browser while the timeline plays.
Use them by hand, or hand them to an agent. Either way, they work on the same timeline.
What I'm not claiming
Video editing isn't solved. An agent doesn't replace someone with taste, and taste is most of the job.
What changes is the middle: finding footage, assembling a first pass, tightening, the fixes nobody wants to do by hand at 11pm. Once you're dealing with hundreds of clips, some of that is tedious enough to be impractical without help.
Try it
If you build with MCP, connect an agent, give it a real timeline, and tell us where it breaks. If you edit for a living, hand it the tedious pass and see what comes back.
Arsaze is live at arsaze.com.