| Repository | Description |
|---|---|
| ai-dj ← you are here | VST3 / AU / Standalone |
| obsidian-neural-gpu-server | Self-hosted GPU server for the 9 engines (Windows, Linux, macOS Apple Silicon) |
| obsidian-neural-controller | Mobile MIDI controller app |
| raveMorph | Neural sound morphing plugin (RAVE-based) |
| beatcrafter | MIDI drum sequencer VST |
Live AI music generation. Standalone or in your DAW (VST3 / AU).
"I've cycled through almost every AI music tool on the market, but Obsidian is the first one that actually feels like a real production tool rather than a novelty. While other AI apps try to replace the songwriter, Obsidian treats AI like a powerful, playable instrument. The 8-track MIDI-triggering is a total game-changer. Because it lives directly in my DAW, there is zero latency and zero break in my workflow. It stays perfectly locked to my project's tempo and vibe, serving as the ultimate intelligent jam partner VST."
- Moteka, Electronic Music Producer SoundCloud · Instagram
The whole plugin - including the local CPU engine - is now free and released under the GNU AGPL v3.0. No license key, no subscription, no credits, no account.
- Stable Audio 3 Medium runs entirely on your own CPU. No GPU, no cloud, no internet required after downloading the model once. Nothing ever leaves your computer.
- The 9 GPU engines run on your own machine too. With obsidian-neural-gpu-server, you host the inference server yourself on an NVIDIA GPU (Windows, Linux) or an Apple Silicon Mac - no account, no API key.
⚡ Runs on a standard CPU. Reference: ~11s per generation on a recent laptop CPU, alongside a full DAW session. 🍎 macOS: Apple Silicon (M1+) only - Intel Macs not supported.
Out of the box, OBSIDIAN Neural generates locally with Stable Audio 3 Medium on your CPU. To unlock the 9 specialized GPU engines, run obsidian-neural-gpu-server on a machine with a supported GPU - the same computer as your DAW, or another one on your network.
| Platform | GPU | Status |
|---|---|---|
| 🪟 Windows 10 / 11 | NVIDIA (CUDA) | Supported |
| 🐧 Linux | NVIDIA (CUDA) | Supported |
| 🍎 macOS | Apple Silicon (M1+, Metal) | Experimental |
- Clone obsidian-neural-gpu-server and follow its README to install it and download the models (
install.baton Windows,install.shon Linux and macOS) - Start the server
- In the plugin: Settings → Server/API → enter the server URL (e.g.
http://localhost:8000) - Pick one of the 9 engines and generate
- Testing and feedback on the Linux and macOS GPU server
- Local GPU inference directly in the plugin (CUDA, Metal, DirectML…)
- More models exported to ONNX for the CPU engine
→ Start a discussion · Open an issue
Type words → Get musical loops. No stopping your creative flow.
- 8-track sampler with MIDI triggering (C3-B3)
- 4 pages per track (A/B/C/D) - Switch variations instantly
- 8 sequences per page - 256 total patterns for complex live sets
- 16-step sequencer with multi-measure support
- Quantized page changes - Seamless transitions locked to measure boundaries
- 4 pair crossfaders with master bypass - Blend each deck A/B pair independently with model-aware color morphing, or toggle the entire crossfader section off for direct routing
- Plug-and-play MIDI mapping - Auto-configured for the companion mobile controller, with bidirectional feedback (LED states, knob positions) on dedicated MIDI channels. To use: open the MIDI panel (piano icon, bottom-right) → Load Default Mapping.
- MIDI learn on every parameter - Map any control to any hardware override, with persistent user mappings
- Per-page ADSR envelope - Shape the dynamics of each variation independently, editable directly on the waveform
- Per-track gain control - Adjust each sample's level (-12 / +12 dB) before mixing, with visual waveform feedback
- Per-track reverse - Instantly flip any page's playback direction for reversed textures and risers
- Per-track transient scatter - Randomize and reposition transients for glitchy, stuttering rhythmic variations
- Per-track glitch sequencer - Sequence glitch effects step by step for rhythmic, stuttering variations
- 8 modulators per track - Eight independent modulation sources on every track to bring movement to your sounds
- Per-track multi-mode filter - LP/HP/BP with 12 or 24 dB slopes, drive, cutoff and resonance for sculpting each voice
- Per-track 8-band graphic EQ + master EQ - Independent frequency shaping on every voice and global bus, from 40 Hz to 15 kHz
- Per-track compressor + master compressor - Full dynamic control on each voice and on the master bus, with threshold, ratio, attack, release and makeup gain
- Per-track limiter + master limiter - Peak protection on every voice and on the master bus with adjustable release and ceiling
- Per-track distortion - 6 character modes (Soft, Hard, Tube, Fold, Diode, Cubic) with pre/post gain and high-pass cutoff
- Per-track bitcrusher - Bit depth and sample rate reduction for lo-fi, digital degradation
- Per-track chorus - Modulation effect with rate, depth, delay, feedback and dry/wet control for added stereo width
- Per-track phaser - Sweeping notch modulation with rate, depth, feedback and stages for movement and width
- Per-track flanger - Classic jet-swoosh modulation with rate, depth, feedback and delay control
- Tempo-synced delay send - 8 time divisions (1/16 to 2 bars), Stereo / Ping-Pong / Mono modes, per-track send level
- Reverb send - Per-track reverb with size, damping, width and mix controls
- Airwindows Console6 master bus - Analog-modeled saturation for cohesive mix glue
- Local CPU generation with Stable Audio 3 Medium - fully offline
- Prompt bank with editor - Build, organize and reuse your prompts with model-aware keywords (genres, elements, moods, negatives)
- Drag-and-drop prompts - Drop a prompt on a track to assign both prompt and AI model in one gesture
- Sample bank with drag-and-drop - Every generation is automatically saved and can be reused across tracks and projects
- Non-blocking generation - No pre-recorded samples, renders in background
The plugin supports 9 specialized AI engines in server mode, powered by your own obsidian-neural-gpu-server - see GPU engines: run your own server.
- stable-audio-open-1.0 - Versatile foundation, drums and full-mix textures (80–160 BPM)
- Stable Audio 3 Medium - Next-gen flexible full tracks, isolated stems, FX (80–160 BPM) (also the local CPU engine)
- Foundation-1 - Tag-based melodic and harmonic phrasing (100–150 BPM)
- Audialab EDM Elements - High-energy EDM leads, supersaws, plucks (100–150 BPM)
- RC Infinite Pianos - Grand and electric piano performances (100–150 BPM)
- RC Vocal Textures - Choral, operatic and atmospheric vocals (100–150 BPM)
- SAO Instrumental - Melodic trap, lofi jazz rap, indie stems (75–160 BPM)
- StableBeaT - Trap beats and 808 grooves (75–160 BPM)
- gluten_v1 - Loopable melodic trap and wavy motifs (90–160 BPM)
⚠️ AI generation can produce unexpected results. Feedback welcome on Issues or Discussions.
- Automatic tempo sync - Incoming samples are BPM-detected with MiniBPM and time-stretched on load via Signalsmith Stretch to lock to your host tempo, with zero pitch drift
- Standalone version - Run OBSIDIAN Neural without a DAW, with built-in transport and tempo control
- Ableton Link (Standalone) - Bidirectional network sync of tempo and start/stop with any Link-enabled app
OBSIDIAN Neural is NOT a song generator like Suno or Udio. It's a performance tool: you build your track loop by loop, you're the composer, AI is your loop generator.
For best performance and sync accuracy with Ableton Link, select "Windows Audio (Exclusive Mode)" in the audio settings, then your audio interface. The default shared mode adds 20-50ms of latency.
If you have ASIO drivers from your audio interface vendor, those will work even better - but require a separate build with the Steinberg ASIO SDK (not redistributable).
- Download the plugin from Releases
- Load it in your DAW (or run the Standalone) → Settings → Local Model
- Accept the model licenses and download the model once
- Type a prompt and generate - fully offline from now on
Want the 9 GPU engines? See GPU engines: run your own server.
New to OBSIDIAN Neural? Read the full walkthrough on KVR Audio: A New Kind of Sampler: Obsidian Neural, Explained by Its Developer
It covers the whole interface step by step: prompt and sample banks, tracks and pages, the step sequencer, mixer channels and crossfaders, per-track effects, the glitch sequencer and its presets, modulators, delay/reverb sends and MIDI mapping.
| Platform | Install path |
|---|---|
| Windows VST3 | C:\Program Files\Common Files\VST3\ |
| Windows Standalone | Run OBSIDIAN-Neural.exe from anywhere |
| macOS VST3 | ~/Library/Audio/Plug-Ins/VST3/ |
| macOS AU | ~/Library/Audio/Plug-Ins/Components/ |
| macOS Standalone | /Applications/ (drag the .app) |
| Linux VST3 | ~/.vst3/ |
| Linux Standalone | Run the binary from anywhere |
Contributions of any size are welcome: bug fixes, new effects, UI improvements, documentation, and especially testing and feedback on the Linux and macOS versions of the GPU server.
💬 GitHub Discussions · 🐛 Issues
🆓 GNU Affero General Public License v3.0 - see LICENSE.
The AI models are not part of this repository and are not covered by the AGPL. They are downloaded separately and remain subject to their own terms: the Stability AI Community License and the Gemma Terms of Use.
Developed by InnerMost47 (Anthony Charretier)
Special thanks to Moteka for the testimonial and early adoption, Stability AI, and all beta testers and early adopters.
Made with 🎵 in France
