Generate images and videos through MuAPI's unified API

MuAPI Media defines an asynchronous image and video workflow that uses the current model catalog and each selected model's schema. It allows model choice without a separate surrounding submit-and-poll integration for every provider. The skill supplies instructions rather than an SDK or background worker, and it does not guess request fields or conceal billable generation.

Model discovery reads the catalog, which supplies names, categories, endpoints, and metadata. The detailed schema is fetched immediately before building a payload to check required fields, types, allowed values, size limits, and outputs. Catalog endpoints already contain the version prefix, so the workflow appends them to the API host without duplicating that prefix.

Preparation reviews the requested media type, model parameters, output destination, and estimated cost. Reference media must use an authorized HTTPS input URL and the exact documented field. The generation POST is submitted once, and its request ID is saved for diagnosis and continuation. A timeout does not justify automatically creating another paid task because the original request may have been accepted.

Use this skill when the user explicitly requests MuAPI image or video generation, or an existing authorized workflow needs a hosted asynchronous media API. It also supports comparing or switching models while keeping the surrounding lifecycle consistent.

Before generation, confirm rights to send the prompt and reference media to a third party, obtain approval immediately before the potentially billable request, and establish whether the request is single or an explicitly approved batch. It is unsuitable for text chat, unchecked provider schemas, or unrelated transfers and bulk generation without authorization.

Inputs include the prompt, requested media type, selected model, schema-confirmed parameters, optional reference URLs, and a controlled output location. MUAPI_API_KEY must be present in the environment or an approved secret manager. The key stays out of chat, request JSON, source files, logs, and process arguments; authenticated headers are passed through protected input or configuration.

Discover available models with a read-only request:

curl --fail --silent --show-error   "https://api.muapi.ai/api/v1/models"   | jq -r '.models[] | [.name, .category, .endpoint] | @tsv'

Outputs include submission and prediction records, a preserved request ID, and a validated media file. Poll /api/v1/predictions/{request_id}/result for the original ID at most 120 times, with a finite deadline and terminal states completed, succeeded, or success; interrupted polling resumes that request rather than submits another. Validation errors require fetching the current schema and rebuilding the payload.

Download an HTTPS output using a fresh request without the API key. Validate redirect destinations and DNS results where needed, then check status, size, content type, and basic image or video decoding. Empty, HTML, or JSON downloads are not usable media merely because HTTP succeeded. Never execute downloaded files as code.

The instructions use MuAPI's catalog, model-schema, generation, and prediction endpoints with curl and jq. Temporary response files need restrictive permissions and cleanup. Model fields, pricing, retention, and availability can change; output URLs may expire, and completed requests still require human review of media quality, rights, and provider restrictions.