claude-music: Generate Full Songs in Your Terminal with ACE-Step 1.5

The Open-Source Model That Beats Suno and Udio on Every Benchmark

The generative AI in music market hit $569.7 million in 2024 and is growing at 30.4% annually toward $2.8 billion by 2030 (Grand View Research, 2024). Suno, Udio, and their cloud competitors are riding that wave - charging $10-20/month for access to models you can't inspect, storing your audio on their servers, and updating their Terms of Service whenever it suits them. Meanwhile, ACE-Step 1.5 - a fully open-source music generation model - outperforms both on every published benchmark, runs on your GPU, and costs nothing.

I wrapped it in a Claude Code skill. claude-music is a 10-command AI music production system that runs entirely in your terminal. Generate full songs from text descriptions. Remix tracks with style transfer. Fine-tune on your own music catalog using LoRA. Export platform-ready files for Spotify, YouTube, or TikTok. Everything stays local. No accounts, no monthly fees, no data leaving your machine.

What Is ACE-Step 1.5 and Why Does It Beat the Paid Tools?

ACE-Step 1.5-XL scores 7.76 on AudioBox, 8.12 on SongEval, 6.62 on Style Align, and 8.42 on Lyric Align - beating Suno v5 (7.69, 7.87, 6.51, 8.29) and Udio v1.5 (7.45, 7.65, 6.15, 8.03) across the board (ACE-Step official, April 2025). It's a diffusion-based model - the same architecture that made image generation explode, applied to audio generation with dedicated music-specific training.

The technical architecture splits into two tiers. The Turbo model is 2B parameters, completes in 8 diffusion steps, and runs on as little as 4GB VRAM - that's an RTX 3060 or better. The XL model is 4B parameters with ~9GB bf16 weights, needs about 16GB VRAM for the highest quality setting, and is what produces the benchmark-leading results above. Both tiers support 10 seconds to 10 minutes of audio, 1,000+ instruments and styles, batch generation of up to 8 songs simultaneously, and 50+ languages for vocals.

According to the ACE-Step team, a full song generates in under 2 seconds on an A100 and under 10 seconds on an RTX 3090 (ACE-Step GitHub, April 2025). That multilingual support matters more than it sounds - most AI music tools struggle with non-English lyrics. If you're creating content for Korean, Japanese, Spanish, or French audiences, ACE-Step 1.5 handles it without the awkward phonetic artifacts you get elsewhere.

10 Sub-Skills: What Can You Actually Do?

84% of developers are using or plan to use AI tools in their development process, but 66% cite AI solutions that are "almost right but not quite" as their biggest frustration (Stack Overflow Developer Survey, 2025). claude-music's 10 sub-skills are designed to give you precise control rather than hoping the model guesses your intent correctly.

Three of these deserve special attention. repaint is what separates claude-music from "generate and hope" - if your chorus is perfect but the intro drags, you describe the timestamp range and what you want changed. That's section-level control, not full regeneration. lora gives you a persistent style fingerprint - train it once, use it on every future generation. export handles the painful loudness normalization step that most creators get wrong: Spotify targets -14 LUFS, YouTube -13 LUFS, TikTok -14 LUFS. One command, correct output.

How Does LoRA Fine-Tuning Work?

ACE-Step 1.5 supports LoRA fine-tuning on 3-10 songs in approximately 1 hour on an RTX 3090 (ACE-Step GitHub, April 2025). This is what makes claude-music genuinely useful for professional use rather than just experimentation. The workflow:

# Step 1: Point it at your training songs
/music lora --train ./my-songs/ --name my-style

# Step 2: Wait ~1 hour (RTX 3090)
# The trainer runs locally, nothing is uploaded

# Step 3: Use your checkpoint in every future generation
/music generate --caption "upbeat summer pop" --lora my-style --duration 60

The 3-10 song requirement isn't arbitrary. Fewer than 3 songs and the model memorizes rather than learns style - every generation sounds too similar to your specific tracks. More than 10 songs and training time extends significantly with diminishing quality returns. The sweet spot for most artists is 5-8 songs that represent the range of your intended output style.

What does this unlock in practice? Podcast producers can train on their existing intro music and auto-generate episode-specific variations that stay on-brand. YouTube creators can build a signature sound for a channel and generate matching background tracks per video. Game developers can train on an environmental audio palette and generate new tracks that fit without manual A/B testing against the existing mix. The key point: your custom style stays on your hardware. Nobody else trains on your music, and no API dependency means no training data policy changes to worry about.

Why Run Music AI Locally Instead of Using Suno or Udio?

86% of global creators now use generative AI in their content work, and 81% say it helps them produce content they couldn't have made otherwise (Adobe Creators' Toolkit Report, October 2025). But cloud tools carry a hidden cost beyond the monthly fee: your output, your style fingerprint, and your training data all live on someone else's server. Music licensing is already complicated - you don't need to add "my cloud provider might have a claim on this" to the legal surface area.

With claude-music, the audio files live on your disk. The model weights live on your disk. The LoRA checkpoints you train live on your disk. Nothing is uploaded during generation or training. The copyright situation is exactly as clear as it would be if you'd produced the audio yourself - which, for AI-generated content, remains an open question in most jurisdictions, but at least it's between you and the relevant law, not between you and a Terms of Service update.

Here's what most coverage of AI music tools misses: the most valuable use case isn't replacing professional musicians. It's giving developers, marketers, and creators access to production-quality background audio without a subscription or a digital audio workstation. The person who benefits most from claude-music isn't a producer looking to automate their workflow - it's the developer who needs three variations of a lo-fi beat for an app demo and doesn't want to spend $40 on a stock audio license, or the marketer who needs 12 regional variations of background music for localized video ads.

How to Install claude-music in 5 Minutes

The installer handles everything: Python environment setup via uv, FFmpeg, and the ACE-Step models (~5GB, downloads with explicit confirmation). You need Claude Code and a GPU with at least 4GB VRAM. Here's the full setup:

# Step 1: Clone the repo
git clone https://github.com/AgriciDaniel/claude-music.git
cd claude-music

# Step 2: Run the installer
bash install.sh           # Linux / macOS
# OR
powershell -ExecutionPolicy Bypass -File .\install.ps1   # Windows

The installer detects your GPU and VRAM, tells you which quality presets are available, downloads ACE-Step with your confirmation, and runs a test generation to verify the setup works. There's nothing to configure manually afterward.

Once installed, open Claude Code and try:

/music generate "chill lo-fi beat, 60 seconds"
/music generate --caption "upbeat pop, female vocal" --duration 90 --quality high
/music random

The 30+ built-in genre recipes handle the prompt engineering for you. You don't need to know that a lo-fi beat needs "boom bap drums, vinyl crackle, rhodes piano, muted guitar" - typing "lo-fi beat" activates the recipe, which supplies the optimal caption structure, BPM range, and instrument weighting automatically.

Frequently Asked Questions

Do I need a powerful GPU to run claude-music?

4GB VRAM is the minimum - that's an RTX 3060 or equivalent. At 4GB you get the Turbo model at standard quality. 8GB unlocks Turbo with extended thinking for better structure. 16GB+ gives you the XL model at full quality - what the benchmark scores above are based on. CPU-only mode works but generation takes 5-10 minutes instead of seconds.

How does the output quality actually compare to Suno or Udio?

On the four published benchmarks, ACE-Step 1.5-XL scores higher than both services across AudioBox, SongEval, Style Align, and Lyric Align (ACE-Step official, April 2025). The XL model's vocal clarity and lyric coherence are the most noticeable improvements over standard quality - particularly in English and for complex song structures with clear verses and choruses.

Can I use the generated music commercially?

The claude-music skill and ACE-Step model weights are MIT licensed. The legal status of AI-generated audio for commercial use varies by jurisdiction and continues to evolve - the same landscape applies here as with any AI content tool. No additional restrictions are added beyond the MIT license and the model's own terms.

What audio formats does the export command produce?

WAV (lossless, for production workflows), MP3 (standard distribution), and platform-specific exports with proper loudness normalization: Spotify (-14 LUFS), YouTube (-13 LUFS), TikTok (-14 LUFS), podcast (-16 LUFS), and CD (-23 LUFS). The /music analyze command can verify the output meets target specs before upload.

Does it work on Windows?

Yes - the PowerShell installer (install.ps1) handles full setup on Windows. The skill works wherever Claude Code runs: CLI, Desktop app (Mac and Windows), and VS Code extension. Developer Mode or Admin privileges are required for the Windows installer to create the necessary symlinks.

Build Your Audio Stack in the Terminal

The AI music market is growing at 30.4% annually and most of that growth flows toward closed, cloud-locked platforms. ACE-Step 1.5 is the first open-source model to match - and in benchmarks, beat - their quality. claude-music puts a full production system on top of it: generate, remix, fine-tune, analyze, and export, all from a single terminal command.

10 sub-skills. 30+ genre recipes. LoRA fine-tuning. Platform export. Zero subscriptions.

• Star the repo on GitHub
• See what else I've built on the about page
• Pair it with claude-canvas for visual + audio production in the same terminal
• Learn more about building tools like this with Skill Forge

Related Posts

• Claude Code Just Turned Obsidian Canvas Into an AI Design Studio - Visual production companion to claude-music
• Best Claude Code Skills in 2026 - The definitive ranking of skills that save time
• Build Your Own Claude Code Skills with Skill Forge - From idea to published skill in one session
• Obsidian AI Second Brain - The knowledge engine that pairs well with the audio stack