Getting Started

Mixture makes video two ways, and they’re the same video.

Visually — a canvas in your browser: design, animate on a timeline, export an MP4. Programmatically — hand a project to a coding agent and watch it build, right alongside you. Nothing is locked to one mode: an agent’s work lands as ordinary layers you can drag, and anything you make by hand an agent can read and edit.

Most people start with the agent, because it’s the fastest way to a finished video — so that’s where this page starts.

Early access, and free

Mixture is in early access and free to use. It moves quickly, and it’s being built in the open with the people using it — including the Supercut pairing, which we’re building alongside the recording side rather than bolting on afterwards.

That means two things worth saying plainly. Some of it is rougher than it will be, and some pieces are still landing. And your feedback genuinely moves the roadmap — the Discord is where early-access users tell us what’s missing, and it’s shaped what shipped this month more than anything else has. If something’s awkward, say so there; it’s a small enough team that you’ll be talking to the people who build it.

What you can make

Hand a project to a coding agent — Claude Code, Codex, Gemini, Cursor, OpenCode — and ask for what you want in plain language.

A cold one-liner works, and it’s the fastest way to see the thing move:

“Make me a 30-second launch video for mixture.dev — pull the copy and colours from the site, and give me a 9:16 cut too.”

Treat that as a starting point, though, not a finished piece. An agent with nothing to work from has to invent everything — and invented brand material looks like invented brand material. The videos people actually ship come from giving it something real to build with:

Put the material in first. Drop your footage, logo, typefaces and music into the project — or into the workspace Library, where every project can reach them. An agent composes from what’s there instead of approximating it, and “use our brand font and the logo” becomes a thing it can actually do.
Point it at a folder. Your agent runs on your machine, so it can see your disk: “the clips are in ~/Recordings/launch — build the story from those, in order.” It uploads what it needs and works from your real takes.
Give it the source, not a summary. A transcript, a CSV of the numbers, the changelog, the script, the brand’s own site. The more concrete the input, the less it has to guess — and data can drive the graphics directly rather than being retyped into them.
Or work the story out together first. If you don’t yet know the shape — what the beats are, what the one message is — say “let’s plan this first.” The agent writes the script and asks you to agree the words, then storyboards it on the canvas: a draft composition of beat cards saying what the viewer sees, hears and should feel in each, plus the look it intends. You redirect that before a single real layer exists. It offers this itself on a thin brief; asking is faster.
Then keep the method. Once one comes out right, save it as a recipe and the next one is a one-line ask with the context baked in.

You don’t need a project open to begin: the agent can create it, build it, and hand you a link. Open that link and you’re watching the canvas fill in as it works, with a live indicator showing which agent is connected and what it’s doing. Or go the other way — open something you’ve started and say “tighten the pacing and add captions.”

Setup is one command, once, and the agent signs in as you. Connect your agent →

What makes it different

Two things worth knowing early, because they’re not how video tools usually work:

Bring your ownYour WGSL shaders, GLB models, typefaces and datasets become first-class assets — and appear as options in the canvas menus, in every project.

Make it once, then every weekKeep a finished video as a recipe and re-run it on new footage, new data, a new language — the treatment holds, the content changes.

Your agent's toolboxVoice from ElevenLabs, video from Runway or Veo, recordings from Supercut — whatever your agent can reach becomes material here. No integrations required.

Fifteen worked examples

Each is a real prompt and what comes back:

Caption a talking headWord-level captions from the audio, styled and timed to the speech.

Turn a website into a promoPoint it at a URL; get a video in that brand's voice.

Screen recording → product demoZooms, cuts and callouts over your raw capture.

Cut to the beatDetected beats drive the cuts and the motion.

Add a voice-overScript, generate, and place narration against the edit.

Podcast → clipsFind the moments, cut them, format them vertical.

All fifteen → — or the same ground area by area: motion, video & audio, captions, data, effects & shaders, 3D, formats and design.

Recording a demo?

If you capture with Supercut, connect both MCP servers to the same agent and a raw session becomes a directed demo — a real cursor driven from the mouse data, punch-ins on the moments that matter, motion blur that matches the move, and a camera you direct. Supercut + Mixture →

Or build it by hand

The canvas is a full editor, not a preview. A project holds compositions — parallel outputs side by side, a 16:9 master next to a 9:16 cut — and each composition sequences scenes in time. Select any layer and the properties panel opens on the right; expand the timeline from the bottom bar to trim clips, place animations, and edit against an audio waveform.

Start with adding layers, then animate one, edit the video and audio, and export.

Make it look good is the one to read after your first video — it’s the difference between “an agent made this” and something you’d publish. And keyboard shortcuts are worth five minutes early.

Where things live

Projects — every video in the workspace you’re in.
Library — the material shared across the workspace: brand fonts, logos, music beds, reusable components. Upload straight into it, or share something from a project (more).
Recipes — a method you repeat, saved: instructions, structure, and the assets that go with it (more).
Community — recipes and components published by other people, ready to start from.

The Library is shared across your workspace, so material is uploaded once rather than per project. Shared team workspaces — inviting people, roles — are not open yet.