chutes-media

Chutes Media

Generate image, video, music, and speech through Chutes and save the asset directly into the project being worked on, then reference it from the code. Use the MCP tools when they are available (list_media_models, describe_media_model, generate_media); otherwise use the chutes-media CLI, which exposes the same three operations and prints a JSON result to stdout.

Golden rule

Model payloads are never uniform across Chutes — FLUX, Qwen-Image, Wan, LTX, ACE-Step and TTS models each expose different fields. Always call describe_media_model before generate_media for any model not already described in this session, and compose the payload from the returned example. Never guess field names.

Core workflow

Follow these steps for every generation request:

1. Discover — Call list_media_models with the target kind (image | video | music | speech) and/or a query. If the user named a model, skip to step 2. Models are not hardcoded and the catalog changes, so discover rather than assume.

2. Describe — Call describe_media_model { model }. Read the returned cords. Each cord lists its required fields, fields (types/defaults/ranges), and a ready-to-fill example; the result also reports a top-level supportsEditing. Pick the cord for the operation (generation cord by default).

3. Compose — Start from the cord's example, set the real prompt/inputs, and keep the payload flat. The server transparently unwraps function-signature schemas (e.g. an input_args wrapper) and validates the payload before spending a GPU call, returning a precise error if a field is wrong.

4. Generate — Call generate_media { model, kind, params, cord?, output_dir?, filename?, overwrite? }. The server warms the model, invokes it (blocking, with progress updates), retries transient cold-start 503s automatically, downloads the asset, and saves it under ./assets/chutes/<kind>/ by default. Outputs must remain inside the workspace; set overwrite: true only when replacement is intentional. It returns { path, kind, model, cord, bytes, contentType, cost?, durationMs, schemaHash?, provenancePath? } — schemaHash pins the cord schema the payload was validated against, and provenancePath is the sidecar written next to the asset (omitted when provenance is disabled).