Skip to content

Agent image, video and audio jobs with ToolRouter

Use ToolRouter for agent tasks that need image generation, video generation, transcription or other media operations exposed to the connected client. Name the output, model preference, input assets and budget before discovery. Not every client profile exposes standalone generation.

Check the client profile first

The Anthropic-specific connector used by Claude chat and Cowork filters standalone AI image, video and synthetic audio generation. It still exposes appropriate screenshot, transcription, analysis and chart operations. Use the tools returned to your client rather than a provider list to determine availability.

Other clients use the general catalog, subject to the chosen operation and account access. Installing a skill does not unlock an excluded category.

Define the task

Example:

Use ToolRouter to discover an image operation for the product scene I describe. Show the selected model, input schema, price and output dimensions before paid work. Use only the reference assets I provide, then return the completed asset URL after checking job status.

For a transcription task, specify the authorised audio file or URL, language and expected output format. For video, define duration, aspect ratio and source assets without inventing unsupported model settings.

Discovery and input assets

json
{
  "query": "generate an image with a specified model and output size"
}

The live catalog includes generate-image operations such as text_to_image, edit_image and list_models, generate-video operations such as text_to_video and image_to_video, and audio-transcriber operations such as transcribe_audio. Providers include fal.ai, OpenAI and ElevenLabs where the chosen operation supports them.

Read the returned schema before supplying inputs. A local filesystem path is not automatically accessible to a hosted provider. Follow files and media to upload or provide a supported file reference, and use only assets you are authorised to share.

Finish the asynchronous job

  1. Check the selected model, price, input limits and approved budget.
  2. Run the discovered operation using use_tool.
  3. If a job handle is returned, follow the advertised status operation, such as job_get.
  4. Continue until output is complete or failure is reported.
  5. Return the final asset links, actual model and relevant output details.

A queued response is not a finished image or video. Do not repeat an uncertain paid generation request before checking its existing job; that can create duplicate work.

Review the result

Open the completed asset and check it against the requested content, format and dimensions. A finished provider job does not establish that the asset fits the brief. Keep generated content distinct from an actual product screenshot or recorded event.

For transcription, preserve uncertain sections and speaker limitations rather than silently rewriting them as facts.

Costs and errors

Models can have different charging units and input limits. Discover the current operation and price, check credits and propose the concrete batch before running it. A provider failure should remain visible in the result; do not substitute a different model or change the budget without authority.

Image tool · Video tool · Transcription · Agent setup

Operations checked in the live catalog on 3 October 2026. Examples describe the workflow and do not claim a paid generation benchmark.