Prompt & Done · CHORUS pack

Claude writes the lyrics. A free AI on your own computer sings them.

ACE-Step 1.5 is an open-source (MIT) music generation model that runs on your own machine. This page has the install steps, the exact Claude request and style line from the reel, the API call, and what we measured on 24 Oct 2026.

♪ Listen: three full 60-second songs

Same lyrics, same style line, same request, re-run on 24 Oct 2026 (three new takes; the reel's 10-second clip is from the first run). Unedited ACE-Step output.

Take 1

Take 2

Take 3

1 Install and start it

# needs Python 3.11 or 3.12 and git
curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv sync
chmod +x start_api_server_macos.sh && ./start_api_server_macos.sh   # Apple Silicon (MLX). On NVIDIA: ./start_api_server.sh

Repo and README: github.com/ACE-Step/ACE-Step-1.5. The first request downloads the models (about 9.4 GB for us), so leave disk space; after that it runs locally. There is also a Gradio web UI (start_gradio_ui_macos.sh) if you prefer clicking to curl.

2 Ask Claude for lyrics

claude -p "Write lyrics for a catchy 60-second pop song about an AI coding agent that fixes my bugs while I sleep. Format for the ACE-Step music model: section tags in square brackets ([Verse 1], [Chorus], [Verse 2], [Chorus]); 4 short lines per verse, 4 short lines per chorus, simple singable words, a repeated hook line 'fixed it while you slept'. Output ONLY the lyrics, no commentary."

This is the exact request we sent (Claude Sonnet, about 5 cents of API-equivalent cost) and the lyrics it returned:

[Verse 1]
Midnight hits, I close my eyes
Terminal glows, the cursor blinks
Red lines waiting, stack traces rise
I let it go, no time to think

[Chorus]
Fixed it while you slept
Every bug is swept
Fixed it while you slept
Wake up, all green, no regrets

[Verse 2]
Morning sun, I check my phone
Tests all pass, the build is clean
Pull request merged, I'm not alone
Best teammate I've ever seen

[Chorus]
Fixed it while you slept
Every bug is swept
Fixed it while you slept
Wake up, all green, no regrets

3 The request that made the song

Save as request.json. Lyrics go in lyrics, one line describing the sound goes in prompt, batch_size: 3 gives three takes in one go.

{
 "prompt": "bright upbeat pop song, clear female lead vocal, punchy drums, synth bass, shimmering guitars, catchy sing-along chorus",
 "lyrics": "<paste the lyrics here>",
 "audio_duration": 60,
 "thinking": true,
 "batch_size": 3,
 "audio_format": "mp3",
 "vocal_language": "en"
}

curl -s -X POST http://127.0.0.1:8001/release_task -H 'Content-Type: application/json' -d @request.json
# returns a task_id; poll POST /query_result with {"task_id_list": ["<task_id>"]}, then download the mp3 from the file path in the result

4 What we measured (24 Oct 2026)

Apple M4 Pro Mac with 24 GB, ACE-Step 1.5 (commit ca1e85f), turbo model with the 5Hz language model, MLX backend.

Test	Result
Three 60-second takes of the lyrics above in one request	87 s total (including the first model load)
Do the takes sing the words? (Whisper large-v3 transcript)	Take 3 matched the lyrics line for line; takes 1 and 2 sang verse 1 correctly and most of the rest with small word swaps. We could only check by transcript, not by ear: listen and keep the best take.
Describe-a-song mode (no lyrics, just a sentence)	Did not follow our brief (it made an unrelated Portuguese song). Write or generate lyrics yourself, as above.

One machine, one set of lyrics. The reel plays a 10-second excerpt of take 3, unedited. We did not compare quality with Suno or other services.

5 Use it responsibly

The README asks you to check that what you generate is original, to disclose AI involvement and to avoid copying protected styles or material. Read the license and disclaimer in the repo before you publish songs commercially.