The engine under a video editor. Bring the GUI, the AI, or both.
SnapCut Core is not a video editor. It is the part underneath one: the timeline, the cuts, audio, effects, titles, rendering and export, with no interface of its own. Everything it can do goes through one API that describes itself, so whatever sits on top is up to you.
That is the bet. AI is making interfaces cheap to build. Ask for a CapCut-style app for vertical clips, a dense Resolve-style editor, or no screen at all, and an AI can put it together in an afternoon, as long as there is a solid engine to put it on. Building that engine is the hard part, so that is what I built.
The same API is how AI connects. An agent can see and measure everything in a project, from transcripts and loudness to silences, scene changes and individual frames, so it edits from facts rather than guesses. AI models come in as plugins, local or cloud. Effects and transitions follow one shader standard, so an AI can write a new one from a description and the engine checks it before it ever renders.
Everything beyond that is a plugin, so the core stays lean and what you can build on it stays open. There is more on that below.
Or use both at once. My own desktop editor is one client, an agent in a terminal is another, and they work on the same project live. A cut the agent makes appears on my timeline as it happens. Every edit records who made it, so I can undo the agent’s changes and keep mine.
Most of the work was deciding where each piece belongs: what stays in the core, what is a plugin, and how to keep the core lean while still useful out of the box. It is written in Rust, and I built it working closely with AI coding agents.
The core
Bring your own GUI
Every app is a client of the same API, over a socket or WebSocket. Build a simple editor, a pro one, or embed editing in your own product.
Bring your own AI
Agents drive it from the command line with a skill file that teaches the edit loop. Models plug in as local or cloud plugins, with no AI baked into the core.
Or both, live
People and agents edit the same project at once. Changes stream to every client, every edit can be a dry run first, and undo knows whose edit is whose.
An AI can see the footage
Transcripts, loudness, levels, pitch, silence and scene detection, exact frames and contact sheets, so an agent can check its own work.
Real editing underneath
Ripple and roll trims, J and L cuts, speed, any number of tracks, keyframes, blend modes, frame- and sample-exact, with a content-keyed frame cache that renders ahead of the playhead.
Words on screen
Titles with presets and animations, any font from a file or Google Fonts, and captions from the transcript with the spoken word highlighted as it is said.
Effects to a standard
Effects and transitions are GPU shaders with a declared set of parameters, validated and test-rendered before use. 125 transitions come as an optional pack.
Plays well with others
Projects are OpenTimelineIO, so an edit opens in Resolve or Premiere. Plugins can be written in any language to connect other tools.
Open source
The engine is MIT or Apache-2.0 and tested on Windows, macOS and Linux.
Plugins
The core does the editing and nothing else. It has no AI models, no voices and no 3D, and it never will. Those live in plugins, kept apart from the engine, so the core stays small and stable while what you can build on it keeps growing.
A plugin is a folder with a manifest and a program in any language. It can add commands, analyzers, generators and packs of effects, run as a background job with progress, and call back into the engine like any other client, with its edits undone as one step. API keys stay in the system keychain, never in the project. Because every editor built on the engine speaks the same API, a plugin written once works in all of them.
That opens the door to almost anything. Speech to text for transcripts, captions and cutting by text. Text to speech for voiceovers, dubbing and fixing a flubbed line. AI green screen, depth maps, upscaling, noise removal and stabilization. Generated b-roll, music and sound effects. A bridge to Blender for 3D, or motion graphics written in HTML. Each one can run on a local model on your own GPU or call a cloud API, chosen per plugin, so privacy, cost and quality stay your call.
Plugins can move footage too, not just change it. One can pull clips from a stock library, a company’s asset store or a cloud drive, and another can send the finished video to YouTube, TikTok or a client portal. Anything with a command line or an API can be wrapped in one.
- Available now
Transcription
Deepgram speech to text, bring your own key. Words land with exact timings and speakers, ready for cutting by text and for captions.
- Available now
125 transitions
The gl-transitions collection as an effect pack. A project keeps a copy of what it uses, so it still renders without the pack installed.
- Possible
Speech to text, local or cloud
Whisper on your own machine for private footage, or a cloud service for speed and speaker labels. Same transcript format either way.
- Possible
Text to speech
ElevenLabs or a local voice model for voiceover, dubbing into other languages, and fixing a flubbed word without a reshoot.
- Possible
AI green screen
Background removal and masking from a segmentation model, with no green screen on set.
- Possible
Depth and maps
Depth maps from AI, so effects know what is near and far: background blur, fog, relighting, text tucked behind a subject.
- Possible
Clean up
Noise and echo removal, upscaling, frame interpolation for slow motion, and stabilization for shaky phone footage.
- Possible
Generated media
B-roll, music and sound effects from generative models, dropped onto the timeline as ordinary clips.
- Possible
Blender and 3D
Send a shot to Blender, render 3D titles, scenes or motion graphics, and bring them back as clips.
- Possible
Motion graphics in HTML
Animated titles, charts and lower thirds written in HTML and CSS, the kind of thing an AI writes well, rendered into video.
- Possible
Footage in and out
Pull from stock libraries, asset stores and cloud drives; publish to YouTube, TikTok or a client portal.
- Possible
Your own
Anything with a command line or an API can become a plugin, in whatever language you already use.