Agent capabilities

Media tools let an agent make pictures, video, speech and films as part of a task.

Generation tools that ask for a capability rather than a model, saved with the task's files and billed at exactly what the provider charges.

What they are

Media tools give a harness's agent tools to make pictures, video clips, speech and music, and to lay clips out on a canvas and export them as a film. The agent asks for what it needs, a picture from text for example, and the result names the model that made it.

Capabilities

Picture from text: A picture from a prompt.
Picture from a picture: A new picture from one you give it.
Video from text: A clip of 1 to 15 seconds from a prompt.
Video from a picture: A clip that starts from a picture.
Speech: Up to 4,000 characters read aloud.
Music: A track of up to five minutes, on your own provider key.
Film assembly: Place clips on a canvas, set a timeline and export a film of up to 40 shots and 10 minutes.

Turning them on

Open the harness's settings and turn on Media. Each capability shows whether it is available and which model serves it. The Videos starter kit comes with media tools on.

Where the results go

Everything a media tool makes is kept with the task, and the canvas is written to the task's files as scene.excalidraw. The renders themselves are kept with the task but are not listed by GET /v1/sessions/{id}/files. Deleting the task deletes its media.

What it costs

Each render is charged once, when it finishes, at the provider's listed price with no markup: per picture, per clip or per second of video, and per 1,000 characters of speech. An attempt that fails but that the provider still charged for is charged too. The Models page lists the per-picture prices.

With your own provider key connected, renders go through it first and your provider bills you directly; those renders count toward neither limit below.

A ceiling per task

Renders on our key are capped at $25 per task and also count toward the task's own cost cap. The $25 is checked before each render starts, so the last render can take a task slightly past it.

Model prices: Every model and its price, pictures billed per image.

Limits

Prompt: Up to 2,000 characters.
One render: Up to 20 minutes and 128 MiB.
Film export: Up to 30 minutes.
Media per task: $25 on our key, checked before each render.

Create API Key

Ready to run this against a live workspace? Keys take under a minute.