Turn raw ideas into finished visual media.

Most media tools specialize too early. A diagram becomes a flattened screenshot, an animation becomes an opaque clip, and a recording loses the evidence that would make its edit easier. Transmute keeps source media, semantic shapes, HTML compositions, timing, and provenance as inspectable inputs to one durable project.

That structure lets an agent work from evidence instead of guessing from pixels. A sketch can become editable geometry, a diagram can become a transparent animated overlay, and a recording can supply motion and timing for the final composition. Raw inputs remain immutable; every transformation is explicit and reproducible.

Compose more than a flattened movie

Multiple displays are first-class video tracks rather than one oversized canvas. Webcam, microphone, and system audio keep independent timing and provenance. Window positions make it possible to zoom into the application that mattered at that moment, even when the recording spans an extended desktop.

Interaction metadata supplies the details viewers feel but editors usually recreate by hand: a stable cursor path, a restrained click highlight, a callout that ends when typing stops, or a camera move that follows one face and then widens for a group.

Give agents structure, not a timeline UI

The command line is the discovery surface. Code mode is the composition surface. A workflow can fan independent analyses out in parallel, join their results at one checked edit transaction, and package the useful sequence as a reusable recipe.

export default defineWorkflow({
  id: "my-demo",
  build(workflow, input) {
    const project = workflow.project.snapshot("project", input.project);
    const evidence = workflow.namespace("evidence");
    const inactivity = evidence.analysis.inactivity("inactivity", { project });
    const edits = workflow.edits.batch("edits", {
      cutRanges: inactivity.select("cuts"),
    });
    return workflow.project.commitEdits("commit", { batch: edits, project });
  },
});

Each operation has a bounded input and output contract. References carry dependencies, immutable receipts carry identity, and planning can show the complete effect and resource envelope before execution. The agent gets ordinary TypeScript for loops, calculations, branching, and reusable functions without turning privileged host actions into ambient authority.

Let polish follow the evidence

The goal is not maximum automation. It is a better first edit with reversible decisions. Transmute can find long intervals with no motion, align independent audio, smooth cursor motion, frame the active window, follow one or many faces, and place image, SVG, GIF, video, or emoji overlays. Manual cuts, speed changes, zooms, pans, and keyframes remain part of the same model.

The same project can carry semantic diagram shapes, local vectorization receipts, generated raster assets, and deterministic HTML. HTML layers can use ordinary images, SVG, Three.js, shaders, or animation libraries against an absolute frame clock, then render with transparency as a standalone loop or as one layer in a larger video.

Keep the local boundary obvious

Raw recordings and local media stay local by default. The open headless core owns inspection, diagram compilation, vectorization, composition, and rendering. A separate hosted control plane provides identity and tightly bounded generated-media operations; it never turns a local workflow into an arbitrary provider proxy.

The public CLI and Agent Skill are released; the desktop editor remains alpha. Their interfaces let new renderers, analyzers, generators, and editors meet at versioned artifact contracts rather than one privileged application. The previous Graphics CLI and API remain compatible while their durable capabilities move into Transmute.