Skip to main content

MCP workflows

Try these prompts after connecting an assistant. The tool list names every tool the server offers. Prompts describe what to ask; the result depends on your account's Addons, media and credits.

Explore without starting a run​

Show me the Addons I can use and their credit prices. Check my available credits. Do not start a run.

The assistant uses list_addons and get_billing_balance. Addon inputs and prices come from the live list; they may differ from this documentation build. Listing them does not start a run.

Create content from an existing image​

Find my product image in the media library. Show me which Addon could turn it into a lifestyle scene, its required inputs and its credit price. Ask me to choose the image and approve the cost before starting.

  1. The assistant finds the image with list_media_assets and reads it with get_media_asset. Check that it selected the right product.
  2. It reads list_addons for the Addon ID, input IDs and current price. An image input uses the asset's { key, mediaType }.
  3. After you approve, it calls start_orchestrator_run with those inputs, maxCredits and a new idempotencyKey for this run.
  4. It follows the returned run with get_orchestrator_run and wait=15 until the status is no longer RUNNING, then shares the result link.

If a run is still running after about ten minutes, the assistant should tell you and let you ask again later. list_running_orchestrator_runs finds your running runs. Follow the Addon tutorials for input preparation and what to check in the result.

Bring an image from your assistant​

Use the product photo I attached to this chat. Import it into my media library and show me its name before starting an Addon.

In ChatGPT, import_chat_file imports an attached file. In a coding assistant that can read the file, the upload sequence is start_media_upload, PUT the bytes to the returned URL with its required headers, then complete_media_upload. If the assistant cannot upload the file itself, show_media_upload opens an upload view in clients that support MCP Apps.

Make and edit an image ad​

Make a social ad for my linen shirt from my media library. Show me ad concepts to pick from first and tell me the price before each start.

  1. The assistant finds the product with list_products (or creates it with create_product), starts ad-concept on it and opens show_choices on that run, which shows the concepts as cards with their previews in clients that show MCP Apps.
  2. After you pick, it starts image-ad with each picked concept: one pick is one run, several are one batch with start_orchestrator_batch.
  3. To change the finished ad's text, position or colour, ask for it. The assistant reads the ad's layers with get_media_asset, starts ad-edit with the changed layers and tells you its warnings.

The API guide shows the same steps as requests.

Recover a run request safely​

If a start request loses its answer, returns a 5xx or reports orchestrator.run_start_in_progress, retry with the same idempotencyKey within an hour of the first attempt. This avoids starting a second run. After an hour, stop and check with the user: that key can start a new run.

An orchestrator.price_above_max_credits error starts nothing. Ask the user before raising the budget or trying again. For invalid inputs or a rate limit, follow the recovery instructions in start_orchestrator_run's description, which the assistant receives with the tool.

Signed result URLs expire after 15 minutes. Keep asset IDs or keys, then read the asset or run again for a fresh link.

Next​