A video editor an agent can drive

Start here

Open the editor and run this in the page. It prints the reference and returns it, so it works as well for a model reading the result as for a person reading the console.

editorAgent.help()             // the whole reference
editorAgent.help('commands')   // commands, state, media, text, program, ui, frames, session, export
editorAgent.describe()         // every command, its parameters, its scope, as data

What a batch looks like

// the document, as JSON
const doc = editorAgent.state()

// find where something was said; the answer is in timeline frames
const hits = editorAgent.findText('pricing')

// one batch is one transaction and one undo unit
editorAgent.dispatch([
  { name: 'clip.split', params: { seq: 'main', track: 0, f: hits[0].startF } },
  { name: 'segment.setTransition', params: { seq: 'main', track: 0, index: 1, kind: 'dip', durF: 12 } },
], { actor: 'agent:me' })

// look at the result the way a person would
await editorAgent.frames(assetHash, 100, 160)   // real frames on one canvas, captioned
editorAgent.program.snapshot()                  // what the monitor is showing

The parts that make it safe to hand over

A batch is a transaction

It applies whole or not at all, and counts as one undo.

Agent edits stay out of the human’s undo

A batch from an agent: actor does not land in the undo stack a person is using.

Sessions

Open one with named scopes, do the work, and revert the whole session in one move.

Checkpoints

Name a point and come back to it.

Dry runs

Ask what a batch would do before it does it.

Introspection

describe() returns every command with its parameters and scope, so the surface does not have to be guessed at.

Derived positions

A segment’s place is the sum of the durations before it. There is no position field to write wrongly.

What it can reach

• The document and every sequence, asset and track, as JSON.
• Import from a File or a URL; transcribe; read the text in the picture.
• Every edit the timeline makes: insert, overwrite, razor, ripple, speed, reverse, loop, nests.
• 34 transitions, 13 entrances, 10 loops, an effect stack, a LUT, chroma key, an anchor and keyframes.
• A per-clip audio chain in one command, and loudness on the way out.
• The programme monitor: seek, play, snapshot, and exact frames rendered for a vision model.
• The activity dock as data, so a long job can be waited on rather than guessed at.
• Export, resolving with the finished run.

Where the documentation is

Three places, all on this origin, all plain text: /AGENTS.md is the reference an agent should read first, /llms.txt is the short pointer, and editorAgent.help(topic) answers in the page itself. The document in the head of every page names the API too, and there is a note in the accessibility tree saying so, because a browsing agent reads that tree rather than the source.

The honest limits

• Export is H.264 video and AAC audio in an MP4, or the sound alone as .m4a. No ProRes, no DNxHD, no H.265 out, no image sequences.
• No team features: no cloud project, no review link, no comments, no shared library. A project lives in one browser, and moves as a file.
• No stock library and nothing generative: no stock clips, no music bed, no AI b-roll, no generative fill.
• No third-party plugins. No OpenFX, no VST, no motion-graphics templates from another app.
• Colour is LUTs and the effect rail, not scopes, curves and secondaries. A .cube grades a clip; there is no vectorscope to grade against.
• Built and checked in Chromium browsers, which is where WebCodecs is furthest along. Other browsers vary by codec and by version.
• It runs in a tab, on one machine. A long 4K timeline is heavier here than in a native application on the same computer.

Questions

Is there an API key?

No. The API is a namespace on the page in the browser that has the editor open. There is no server and nothing to authenticate against.

Can it run headless?

It runs wherever a browser runs, including one driven by CDP or Playwright. It needs a real browser because the decoding and encoding are the browser’s.

Can an agent undo its own mess?

Yes, through a checkpoint or a session, which is why agent batches are kept out of the human’s undo stack.

Is there an MCP server?

Not yet. The surface is a documented JavaScript API on the page; anything that can evaluate JavaScript in a tab can drive it today.

Where do I read the full command list?

At /AGENTS.md on this origin, or by calling editorAgent.describe() in the page.