🧄 aioli

Your at-home AI studio: clone voices, compose songs, draw images,
generate video and produce podcasts. All on your machine. No cloud.

🎙️ Voice  •  🎵 Music  •  🖼️ Image  •  🎬 Video  •  🎧 Master  •  📻 Podcast

What it does (with real examples)

19 songs, their cover art and several clips were made this way, locally, with these commands. No accounts, no subscription.

How it works

1. aioli install inspects your machine and tells you what fits your GPU.
2. aioli models download fetches verified weights (sha256, resumable) into ~/.local/share/aioli (or wherever AIOLI_MODELS_DIR points).
3. You generate locally: not a byte leaves your machine. No accounts, no subscription. Only licensed weights (Flux, LTX) ask for a free HuggingFace account to accept their license once.

Install

curl -fsSL https://aioli.es/install.sh | sh   # installs uv+ffmpeg if missing → doctor → setup → status (setup = vía compat del instalador; equivale a `aioli install`)
# …or truly by hand (needs `uv`, Python ≥3.12 and `ffmpeg`) — two commands, no extras:
uv tool install aioli-cli            # the CLI only (base dist, sin extras)
aioli setup                          # usage screen: tick what you want; the assistant installs piper/kokoro and downloads only that
aioli doctor                         # preflight: CUDA, ffmpeg, models, VRAM
# advanced alternative: `uv tool install "aioli-cli[voice-piper,music-storage]"` still works; torch worlds stay manual (`--with`, see plugins/voice/README.md)

For top-tier image and video (Flux, LTX), accept their license on huggingface.co and export HF_TOKEN (read-only token) before install.

🔒 Paranoid (good): the installer lives in the repo's install.sh. It only uses the network for astral.sh (uv), GitHub and HuggingFace. The only elevated step is installing ffmpeg if missing — it asks first (or skips with --no-ffmpeg). Re-running is safe: it upgrades in place and preserves your extras and --with engines.

Requirements

aioli install detects your machine (tier patata/modesta/capaz/bicha — yes, potato to beast) and only downloads what fits. Weights are sha256 verified and resume if your connection drops.

Usage

Camino feliz: aioli install → aioli <plugin> "texto" → aioli doctor (solo si algo falla). Per-model recipes live in docs/USAGE.md.

# image: text → PNG (editing, inpainting and reference available)
aioli image "a lighthouse in the fog"
aioli image "a lighthouse in the fog" -m z-image-turbo -o faro.png

# music: style + your lyrics → full song
aioli music "cheerful rumba" -l my-lyrics.txt -o track.mp3
aioli music generate --preset reggae-consciente -l my-lyrics.txt -o track.mp3

# voice: create profile, add a sample with its transcript, generate
aioli voice profiles create vega --language es
aioli voice profiles add-sample vega ref.wav -t "exact audio transcript"
aioli voice "Hola mundo" --profile vega --engine chatterbox -o hola.wav

# video: text or image → clip with audio
aioli video "a robot garlic dancing" --frames 81 -o ajo.mp4

# podcast: script with [VOICES], [MUSIC:] and [PAUSE:] → episode + notes
aioli podcast episode script.md -o ./ep1

# master: finish any audio (mix, ducking, -16 LUFS…)
aioli master loudnorm mix.wav -o final.wav

aioli <plugin> --help details every command; aioli doctor diagnoses failures. Note: the CLI itself speaks Spanish — the models don't mind.

License

MIT. Third-party weights carry their own license (some non-commercial or requiring acceptance on HuggingFace); aioli models warns you.