AI · new in 3.0
AI agents
The Agents tab is a co-designer that works inside your project. Ask for a title card, a lyric video, a tighter edit, a voiceover or a grade — it builds it layer by layer on your canvas, looks at what it made, fixes what it sees, and hands it back as a single undo. It runs on any model you choose.
Start in a minute
- Open Agents in the left rail (under Plugins).
- Connect a model — pick a provider, paste its API key, press Test. No key? Run a free local model with Ollama. See Connect any model.
- Type what you want. Vague is fine; exact is fine.
Prefer Claude Code, Claude Desktop or Cursor? They can drive AtomCut directly — see Claude Code & external agents.
What it can do
Everything you can do with a pointer, it can do through AtomCut's command API — the same one plugins and external agents use. The API describes itself from the app's live schemas, so a feature that ships tomorrow is usable by the agent the day it lands.
| Ask for | What happens |
|---|---|
| Design | Layouts on a grid, a type scale, colour roles — title cards, lower thirds, posters, promos, carousels. Auto layout, components, nested comps. |
| Animate | Entrances and exits, keyframes with real easing, staggered text, motion paths, path morphs, expressions, draw-on strokes, transitions. |
| Type & captions | Word-synced captions from speech, karaoke and Spotify-style lyric scrolls, kinetic type, text effects and templates. |
| Edit video | Remove pauses and dead air, cut to the best lines, J- and L-cuts, speed ramps, punch-ins, long → short reels, keeping captions in sync. |
| Colour | Grades built from a feeling ("warm, filmic, golden hour") on the node-based grade graph; looks, keys, adjustment layers. |
| Sound | Sounds fired by animation, voiceovers, sound effects and music (with a media provider), ducking, clean-up. |
| Make media | Draw and edit images, shoot and edit video, cut out subjects, upscale — filed in Media with the prompt and model that made them. |
| Study references | Attach an image or a video: it reads the layout, type, colour and motion, and rebuilds the idea with AtomCut's own tools. |
It looks at its own work
After a visible change the agent renders the frame it changed — or a contact sheet of the whole animation — and judges it the way you would: overlaps, cramped margins, weak contrast, jumpy easing. The renders go through the exporter, so what it sees is what you will export.
You stay in control
- One undo per request. Everything the agent does for one message is one ⌘Z, labelled AI · …. Each reply that changed the project has Undo changes under it — it rewinds exactly that request, even from further back, and Redo changes brings it back.
- Steer while it works. Anything you send mid-run joins at its next step. Esc stops it; a stop leaves a clean project.
- See everything. Its thinking, every step it takes, what it looked at, how long it took and the tokens it used are shown live. Click a step to see exactly what it sent and got back.
- Switch models mid-chat from the model menu under the message box. The same menu sets effort: Fast, Balanced, Deep, Max. A ring shows how full the context is.
- Resilient. A busy provider is retried (up to ten times, with the wait shown); a bad key or a refusal says so plainly.
Files, voice and the canvas
- Paste, drop or attach any file — images it can see, video it can study frame by frame, sounds, fonts it can install, Markdown, JSON and subtitles it can read.
- 🎤 dictates into the message box.
- Right-click any layer ▸ Ask AI…, or ▸ AI actions for one-click jobs — see AI actions & Generate.
Plans and privacy
AI agents are part of AtomCut Pro; setup, history and skills are browsable on every plan. You bring your own model: keys stay in your browser and are sent only to the provider they belong to. Your project never leaves your device except what the model needs to see for the request you made.