For Power Users & Local AI Enthusiasts
The local AI music studio: generate, edit, mix, and export full songs, entirely on your GPU.
Deterministic seeds. 120+ CLI commands. No telemetry. No cloud.
GGUF Optimized Deterministic Fully Automatable
Cloud tools see every prompt, every lyric, every idea you type. Foundry's generation is 100% local. Prompts and audio never leave your PC.
Per-generation credits punish experimentation. You have the hardware, so why should a cloud service meter your own GPU? Foundry generates unlimited.
Most cloud tools don't expose seeds or give different results on repeat. Foundry is fully deterministic: same seed, same prompt, same output. Every time.
Web UIs don't pipeline. You can't automate 50 generations from a shell script. Foundry's 120+ CLI commands expose every operation, locally or over network.
Click any video to watch real operations inside Demodokos Foundry.
Text prompt to full song. GGUF inference, 5Hz language model, DiT diffusion. All local.
Natural language to structured metadata. Qwen3-based GGUF with adjustable reasoning.
120+ commands. Local or network operation. Scriptable, deterministic, reproducible.
Up to 7 stems: vocals, drums, bass, guitar, piano, strings, other. Built-in AI separation.
Optimized GGUF models (Qwen3-based, 4B to 35B parameters) with CUDA 12/13 inference. Converts natural language instructions into full song metadata: caption fields, structured lyrics, genre/instrument/vocal config, all guided by a classification and music knowledge pipeline.
A 5Hz language model generates audio code tokens. GBNF constrained decoding ensures structured output. A secondary negative-prompt correction model reprocesses the codes (similar to CFG). Five DiT variants (Turbo, Turbo Detail, Turbo Clean, Turbo Continuous, Base) handle diffusion at different speed/quality tradeoffs.
The GGUF-quantized 5Hz models run at up to 250 tok/s on a 5090 (vs 20 tok/s with PyTorch). Critical for long-form audio: a 10-minute track has ~50,000 code tokens. Ultra-VRAM saver cycles models for 6 GB cards.
Patch (region regeneration with spectral crossfade), Extend (continuation + boundary blending), Cover (audio restyling with adjustable strength), Finish (compositional closure), and Separation (up to 7 stem tracks via AI source separation).
33 effects in 7 groups (Filters/EQ, Time/Space, Dynamics, Stereo, Pitch/Voice, Vintage, Creative/Glitch). 200+ presets. Non-destructive stacking. Full multi-track timeline with trim, fade, speed (0.25x–4x), volume, merge, split, undo/redo.
120+ commands. Operates locally or over network. Exposes the entire application state: create, generate, compose, stem-separate, timeline operations, export. Built-in help with examples. Your AI agent can control it end-to-end.
# Compose with Creative AI (local GGUF inference)
$ foundry creative-agent-process "synthwave, neon city, 120 seconds"
Creative AI composing… (Qwen3 35B-A3B, reasoning: Medium)
# Generate with deterministic seed
$ foundry generate --seed 1337 --variations 4
Generated 4 variations (seeds 1337-1340)
# Patch a rough region (surgical regeneration)
$ foundry patch --start 12.5 --end 15.0 --seed 42
Patched 2.5s region, spectral crossfade applied
# Stem separation
$ foundry stem-separate
7 stems: vocals, drums, bass, guitar, piano, strings, other
# Apply DSP effect chain
$ foundry dsp-apply --preset "Cathedral" --track 1
Applied Cathedral reverb (non-destructive)
# Export everything
$ foundry timeline-export neon_city.flac --per-track
Exported 7 tracks → neon_city/Network-capable: control Foundry remotely from another machine or your own AI agent.
Full seed control. Same seed + same config = bit-identical output. Version-control your generation parameters alongside your project.
Ships its own Python environment. Models download automatically on first run. No PATH pollution, no pip install, no conda. Just run the installer.
Cycles models in and out of VRAM automatically. A 6 GB card runs the full pipeline at the same quality, just slower. 32 GB+ enables extreme throughput.
No usage analytics. No prompt logging. Internet is required only for initial license activation. Generation, editing, and export are fully offline.
The control utility supports network operation. Run the studio on a beefy workstation, control it from your laptop, or let an AI agent drive it entirely.
Export full mix, selection, or individual tracks. FLAC on Pro tier. Batch export from CLI and pipe into your DAW or deployment script.
Every track above was created by Foundry from a simple text description. No pre-made music, no recording studio needed.
Start free, then upgrade when you're ready. All plans run entirely on your computer, no cloud processing, no waiting, no limits.
Start creating today
Full creative suite
API + Automation
All processing runs on your computer. Private, fast, no internet needed.
Cancel anytime, no further charges. Try the Free tier first to explore.
The pipeline adapts to your VRAM. More VRAM = everything loads faster and stays resident.
Windows 10 or 11, 64-bit
macOS coming soon
SSD/NVME disk recommended
NVIDIA GPU with 6 GB+ VRAM
Any RTX series · CUDA 12/13
6–8 GB: Ultra-VRAM saver (cycles models)
16 GB+: multilingual Creative AI resident
24 GB+: full model residency
32 GB+: extreme throughput
Download Demodokos Foundry and start creating. 100% local. No cloud. Just you and your music.
C474B56A67CC42BF11FCB7770808CF8F6ADD9ACCD77953876EF49DFC92CF17DB
macOS coming soon · 25% off all paid plans