The music_generate tool creates music or audio through the shared music-generation capability, backed by ComfyUI, fal, Google, MiniMax, and OpenRouter.

For session-backed agent runs, music_generate starts as a background task, tracks progress in its media runtime, then wakes the agent when the track is ready so it can tell the user and attach the finished audio. The completion agent follows the session's visible-reply contract: automatic final reply when configured, or message(action="send") when the session requires the message tool. If the requester session is inactive or its wake fails and generated audio is still missing from the reply, Apus sends an idempotent direct fallback with just the missing audio.

Quick start

    Set an API key for at least one provider — for example
    `GEMINI_API_KEY` or `MINIMAX_API_KEY`.

    ```json5
    {
      agents: {
        defaults: {
          mediaModels: {
            music: {
              primary: "google/lyria-3-clip-preview",
            },
          },
        },
      },
    }
    ```

    _"Generate an upbeat synthpop track about a night drive through a
    neon city."_

    The agent calls `music_generate` automatically. No tool
    allow-listing needed.

Without a session-backed agent run (direct/local contexts), the tool
runs inline and returns the final media path in the same tool result.

    Configure `plugins.entries.comfy.config.music` with a workflow
    JSON and prompt/output nodes.

    For Comfy Cloud, set `COMFY_API_KEY` or `COMFY_CLOUD_API_KEY`.

    ```text
    /tool music_generate prompt="Warm ambient synth loop with soft tape texture"
    ```

Example prompts:

Generate a cinematic piano track with soft strings and no vocals.

Generate an energetic chiptune loop about launching a rocket at sunrise.

Use action: "list" to inspect available providers/models, and action: "status" to inspect the active session-backed music task:

/tool music_generate action=list
/tool music_generate action=status

Direct generation example:

/tool music_generate prompt="Dreamy lo-fi hip hop with vinyl texture and gentle rain" instrumental=true

Supported providers

MiniMax registers two provider ids sharing the same models: minimax for API-key auth and minimax-portal for OAuth. Model refs follow the auth path (minimax/music-2.6 vs minimax-portal/music-2.6); see MiniMax.

fal also exposes fal-ai/ace-step/prompt-to-audio (wav, no lyrics, no instrumental toggle) and fal-ai/stable-audio-25/text-to-audio (wav, prompt-only) alongside its default MiniMax-backed model. Google's default lyria-3-clip-preview outputs mp3 only; lyria-3-pro-preview also supports wav. MiniMax also exposes music-2.6-free, music-cover, and music-cover-free. OpenRouter also exposes google/lyria-3-clip-preview.

Capability matrix

The explicit mode contract used by music_generate, contract tests, and the shared live sweep:

Tool parameters

Provider request timeouts are operator configuration only. Apus uses agents.defaults.mediaModels.music.timeoutMs when configured, raises values below 120000ms to 120000ms, and otherwise defaults provider requests to 300000ms.

Async behavior

Session-backed music generation runs as a background task:

• Background task: music_generate creates a background task, returns a started/task response immediately, and posts the finished track later in a follow-up agent message.
• Duplicate prevention: while a task is queued or running, later music_generate calls in the same chat return task status instead of starting another generation. Use action: "status" to check explicitly. A recently completed matching request is also deduplicated for 2 minutes. Direct chats keep separate tasks even when they share the main session transcript; completion returns to the requesting peer.
• Status lookup: use music_generate with action: "status".
• Completion wake: Apus injects an internal completion event back into the same session so the model can write the user-facing follow-up itself.
• Prompt hint: later user/manual turns in the same session get a small runtime hint when a music task is already in flight, so the model does not blindly call music_generate again.
• No-session fallback: direct/local contexts without a real agent session run inline and return the final audio result in the same turn.

Task lifecycle

The media runtime reports generation progress:

Configuration

Model selection

{
  agents: {
    defaults: {
      mediaModels: {
        music: {
          primary: "google/lyria-3-clip-preview",
          fallbacks: ["fal/fal-ai/minimax-music/v2.6", "minimax/music-2.6"],
        },
      },
    },
  },
}

Provider selection order

For music_generate, Apus tries providers in this order:

1. model parameter from the tool call. When set, only this model is tried.
2. agents.defaults.mediaModels.music.primary from config.
3. agents.defaults.mediaModels.music.fallbacks in order.
4. When neither a primary nor fallback model is configured, auto-detection using configured provider defaults:
   • current default text-model provider first, if it also offers music generation;
   • remaining registered music-generation providers, alphabetically by provider id.

If a provider fails, the next candidate is tried automatically. If all fail, the error includes details from each attempt. Each failed candidate logs its provider, model, and error at warn. For reference-image requests, candidates that cannot use images or accept the supplied reference count are skipped.

Explicit music model configuration limits fallback to the configured list; Apus does not append auto-detected providers.

Provider notes

Workflow-driven and depends on the configured graph plus node mapping
for prompt/output fields. The `comfy` plugin plugs into the shared
`music_generate` tool through the music-generation provider registry.

Uses fal model endpoints through the shared provider auth path. The
bundled provider defaults to `fal-ai/minimax-music/v2.6` and also exposes
`fal-ai/ace-step/prompt-to-audio` and
`fal-ai/stable-audio-25/text-to-audio` for prompt-to-audio requests.
Lyrics and instrumental mode are MiniMax-model-only; the other two
models are prompt-only.

Uses Lyria 3 batch generation. The current bundled flow supports
prompt, optional lyrics text, and optional reference images. The
default `lyria-3-clip-preview` model outputs mp3 only; the
`lyria-3-pro-preview` model also supports wav.

Uses the batch `music_generation` endpoint. Supports prompt, optional
lyrics, instrumental mode, and mp3 output through either `minimax`
API-key auth or `minimax-portal` OAuth. Also exposes `music-2.6-free`,
`music-cover`, and `music-cover-free` models.

Uses OpenRouter chat completions audio output with streaming enabled. The
bundled provider defaults to `google/lyria-3-pro-preview` and also exposes
`openrouter/google/lyria-3-clip-preview`.

Choosing the right path

• Shared provider-backed when you want model selection, provider failover, and the built-in async task/status flow.
• Plugin path (ComfyUI) when you need a custom workflow graph or a provider that is not part of the shared bundled music capability.

If you are debugging ComfyUI-specific behavior, see ComfyUI. If you are debugging shared provider behavior, start with fal, Google (Gemini), MiniMax, or OpenRouter.

Provider capability modes

The shared music-generation contract supports explicit mode declarations:

• generate for prompt-only generation.
• edit when the request includes one or more reference images.

New provider implementations should prefer explicit mode blocks:

capabilities: {
  generate: {
    maxTracks: 1,
    supportsLyrics: true,
    supportsFormat: true,
  },
  edit: {
    enabled: true,
    maxTracks: 1,
    maxInputImages: 1,
    supportsFormat: true,
  },
}

Legacy flat fields such as maxInputImages, supportsLyrics, and supportsFormat are not enough to advertise edit support. Providers should declare generate and edit explicitly so live tests, contract tests, and the shared music_generate tool can validate mode support deterministically.

Live tests

Opt-in live coverage for the shared bundled providers (fal, Google, MiniMax, OpenRouter):

APUS_LIVE_TEST=1 pnpm test:live -- extensions/music-generation-providers.live.test.ts

Equivalent repo wrapper, which drives the same test file:

pnpm test:live:media:music

This live file uses already-exported provider env vars ahead of stored auth profiles by default, and runs both generate and declared edit coverage when the provider enables edit mode. Coverage today:

• google: generate plus edit
• fal: generate only
• minimax: generate only
• openrouter: generate plus edit
• comfy: separate Comfy live coverage, not the shared provider sweep

Opt-in live coverage for the bundled ComfyUI music path:

APUS_LIVE_TEST=1 COMFY_LIVE_TEST=1 pnpm test:live -- extensions/comfy/comfy.live.test.ts

The Comfy live file also covers comfy image and video workflows when those sections are configured.

Related

• ComfyUI
• Configuration reference — agents.defaults.mediaModels.music config
• Google (Gemini)
• MiniMax
• Models — model configuration and failover
• Tools overview
• Media overview — how the media tools fit together