Keytween

Get started

Overview

Keytween generates images and short videos with sound, paid per generation by GPU time. Everything the web app does is also available to your code and to AI agents:

  • MCP server at https://keytween.com/mcp for agent harnesses such as Claude Code, Cursor, Codex and Gemini CLI. One config block and your agent can generate, edit, wait for results, look at them, quote prices and clean up. Start with Connect your agent.
  • REST API at https://keytween.com/api/v1 for scripts and services. See the REST API reference.

Both use the same API key, spend the same credits and work on the same data, so anything made through one shows up in the other and in the web app.

What you can generate #

You want Use Typical cost
An image from a description Text to image 1 credit at 0.5 MP, 2 ($0.02) at 1 MP, 7 at 4 MP
An image that includes a specific character, product or style Text to image with @references a little more per reference
A changed version of an image Image edit 2 credits at 1 MP, 5 at 2 MP
A video from a description alone Text to video 10 credits ($0.10) for 4.33 s
A video that brings an image to life Image to video same
A video that goes from one exact image to another First and last frame to video same
A longer film made of several shots that join seamlessly A sequence of steps, then export sum of its parts; export is free

Videos are 3.33 to 10 seconds per clip, 24 or 48 frames per second, with generated sound, speech and music. Standalone images go up to 4 megapixels; video frames range from 0.3 to 1 megapixels; any aspect ratio from 9:21 to 21:9 works. See Sizes and durations.

Concepts #

Image library. Where standalone images and their edits live, up to 4 megapixels, with its own references. Any image can later be animated by starting a sequence from it.

Sequence. The container for video: a single clip is a sequence with one step, a multi-shot film is a sequence with many. (The REST API calls it a project.)

Step. One shot. A step has a start frame, an optional end frame and one or more takes. (The REST API calls it a segment.)

Start frame and end frame. The images a shot begins and finishes on. You generate a start frame from text or upload one; you make an end frame by editing the start frame into the state the shot should end in. A step can hold several candidates of each; one is selected.

Take. A rendered video clip for a step. Render as many as you like and select the best.

Accepting a step. Locks the selected take and opens the next step, whose start frame is the take's exact last frame. That is what makes consecutive shots join without a visible cut.

Reference. A supporting image with an @label. Mention the label in an image or edit prompt to carry that character, object or style into the result.

Export. One MP4 stitched from the selected takes of every step, with sound.

Job. Every generation runs as a queued job with a status, progress and an estimated time. See Jobs, waiting and limits.

Credits. Prepaid; 1 credit is $0.01. A job is charged when it is accepted and refunded automatically if it fails or is cancelled. See Credits and pricing.

The three video modes #

Mode Input Best for
Text to video A prompt Fast ideas, b-roll, shots where the exact look is not fixed in advance
Image to video A start frame and a prompt Animating a specific picture: a product shot, a portrait, an illustration, a scene you already designed. The ending is up to the model
First and last frame A start frame, an end frame and a prompt Exact outcomes: transformations, before-and-after, time-lapses, a camera move with a defined destination, loops, and shots that must hand over to the next one precisely

All three produce sound and cost the same.

Where to go next #

All docs