---
name: guaardvark-music
description: "Generate full songs with vocals or instrumentals (ACE-Step) and sound effects or ambience (Stable Audio Open) on the user's GPU through Guaardvark's Audio Foundry. Use when the user asks for a song, a beat, a jingle, background music, or a sound effect, made locally with no subscription."
license: "MIT"
metadata:
  decimalai-skill-id: "b6abb7c4-5a37-4e81-ada1-7cfcc94a0334"
  decimalai-version-id: "f5757bdd-4caf-463c-a9cd-627c13875935"
  decimalai-version: "2"
  decimalai-body-sha256: "c59ba6f2a93d7e16fba5d60685cb6ee4d70ee612448b9a5cd976fd288b7a3641"
  decimalai-name: "guaardvark/music"
  decimalai-format: "1"
  decimalai-invocation: "model"
  decimalai-category: "design-media"
  decimalai-source: "https://app.decimal.ai/s/guaardvark-music@2/SKILL.md"
---
# Music and sound effects with Guaardvark

Read `setup` first; needs the `audio_foundry` plugin. `B=${GUAARDVARK_URL:-http://localhost:5000}`.
ACE-Step takes ~10 GB VRAM; the orchestrator evicts other models while it runs.

## A song

Over MCP, call `generate_music` (`style`, optional `lyrics`, `seconds` up to 240,
`instrumental`, `seed`, `model`); it returns a `job_id` for `get_generation_status`, which reports
the file and a download link when the song is done. Without MCP, use REST:

```bash
curl -s -X POST $B/api/audio-foundry/generate/music -H 'Content-Type: application/json' -d '{
  "style_prompt": "indie folk, fingerpicked acoustic guitar, warm male vocal, 96 bpm, intimate",
  "lyrics": "[verse]\nLines here\n[chorus]\nHook here",
  "negative_prompt": "distorted, muddy, synth",
  "duration_s": 90, "instrumental_only": false,
  "output_format": "wav", "seed": 1234, "async": true
}'
```
- `model`: omit (or `"ace-step"`) for ACE-Step v1; `"ace-step-1.5"` for ACE-Step 1.5 (MIT,
  48 kHz, turbo model plus a planner) once it is installed in Audio Studio → Manage models.
  Not installed → `400` with `needs_install`; nothing is downloaded.
- `style_prompt` uses ACE-Step's tag vocabulary: genre, instruments, mood, tempo, vocal type.
  Vague words ("professional", "futuristic") drift to the model's prior; be concrete.
  `negative_prompt` is accepted but neither ACE-Step model applies it (the result says
  `negative_prompt_applied: false`). `POST $B/api/audio-foundry/rewrite-music-prompt`
  turns plain English into the tag form when the user wants help.
- `lyrics` with `[verse]` / `[chorus]` / `[bridge]` markers; omit or set `instrumental_only: true`.
- `duration_s` up to 240. Always send `"async": true` for anything over a few seconds.
- Response `202 {"job_id"}` → poll `GET $B/api/audio-foundry/jobs/<job_id>`; done when `status`
  is `done`, with `path`, `document_id`, `duration_s`, `seed`. Reuse the seed to reproduce a take.

## A sound effect or ambience

```bash
curl -s -X POST $B/api/audio-foundry/generate/fx -H 'Content-Type: application/json' -d '{
  "prompt": "rain on a tin roof, distant thunder, no music", "duration_s": 20, "output_format": "wav", "async": true
}'
```
`duration_s` up to 47. Same job polling.

## Then

- A finished song can go straight into the music skill-video as the `song`.
- Jobs: `GET $B/api/audio-foundry/jobs` lists, `DELETE` clears finished ones.
- Attribution comes back in the job result; keep it with the file if the user publishes.