A machine your agent edits video on.

What a box gives your agent

Eleven verbs. Each prints a line or two, because every word an agent reads costs it tokens.

roxy new

A machine with a work folder, ready in milliseconds. FFmpeg, Node, Bun, Python and headless Chrome are on it, so Remotion and browser renders run as they are.

roxy put

Upload footage, files or whole folders, or hand it a link and the server fetches the footage itself. Stored by content, so nothing goes up twice. Big files travel in parts, resume after a break, and are usable about a second in, while the rest uploads. Originals are read-only.

roxy exec

Run any command. Back comes the exit code and the last lines, not ten thousand lines of encoder log. The full log stays in the box.

roxy ssh

A real terminal inside the box, from your shell or from the dashboard, for when a person wants to look around.

roxy look

Frames from any clip or finished video as one small image. How an agent picks a crop and checks its own cut without downloading anything.

roxy hear

What is said in a clip and when, a line per phrase, with every word's timing saved in the box. It also measures how loud the sound is around any moment, so an agent that cannot listen can still cut in the quiet between two words.

roxy push

Send a project folder: scripts, a Remotion project, fonts. In a git repository only what git does not ignore.

roxy fork

A copy of the box to try another cut on. Footage is shared, not copied, so forking an hour of 4K takes no time.

roxy review

A page for the cut instead of a file. Whoever has the address watches it, pins notes to exact frames and marks the picture. The next cut appears at the same address.

roxy notes

Those notes, for the agent: the second, the words, where on the picture, and the marked frames as one image. It can wait until the reviewer presses Send.

roxy get, link

Bring a file or a folder home, or hand someone an address for it.

Short form to 4K, delivered right

roxy export cut.mp4 --for instagram checks a finished file against where it is going and re-encodes only what is technically off. It never crops or pads: the shape of the picture is your agent's call.

Three ways in

The same verbs everywhere, so what your agent learns in one place holds in the others.

roxy new
roxy put clip.mp4
roxy exec 'ffmpeg …'

claude mcp add roxy -- roxy mcp
codex mcp add roxy -- roxy mcp

roxy open

Watch it work

Every box has a page. Footage appears as it uploads, commands scroll as your agent runs them, and each new cut plays the moment it exists.