
Mix in your DAW
Already working in Pro Tools, Logic or Ableton? Stay there. Put the StemDesk Channel Strip on each track: the app works out what each track is and sets the plugin up for it. Change anything in the app and your DAW follows.
Thank you for your purchase! Your license key is on its way from licenses@stemdesk.ai (allow a couple of minutes). It often lands in Spam/Junk — if it’s there, please mark it Not Spam. Nothing after 15 minutes? Email support@stemdesk.ai.
Skip to content A mixing app +
17 plugins, linked
StemDesk is a mixing app plus 17 plugins that work together. Mix right inside your DAW, bring in your stems, or make a new song from scratch. Your AI engineer mixes and masters it for you, and shows you every move.
Free tier includes the Compressor, Gate and Limiter plugins — no account required. Pro $99 · Studio $179, one-time, no subscription. See pricing
Desktop app · 17 plugins · local AI engineer · any VST3, AU or AAX DAW
How it works
Start wherever you are. Your AI engineer handles the rest.

Already working in Pro Tools, Logic or Ableton? Stay there. Put the StemDesk Channel Strip on each track: the app works out what each track is and sets the plugin up for it. Change anything in the app and your DAW follows.

Exported your tracks? Bring the stems folder or a Pro Tools session into the app. Only have the finished song? StemDesk splits it into vocals, drums, bass and the rest.

No song yet? Describe your idea. The AI writes the lyrics, then produces the whole track, with beat, instruments and sung vocals, and hands it to you as separate stems.

Your AI engineer balances the levels, sets EQ, compression and panning, and masters the song. Every move comes with the reason, and you can change any of it.

Export WAV, MP3, FLAC or AAC, ready for Spotify or YouTube. Or send the whole mix back into your DAW as a session.
Music generation is an experimental feature you switch on in Settings → Labs; it uses ACE-Step, a free separate install you point StemDesk at. Lyrics, separation, mixing, mastering and everything else work without it.
Drop in a stereo 2-track and StemDesk splits it into 4 or 6 stems, locally. Or start from your own stems folder, a Pro Tools session, or a reference track.
Every stem is measured — role, LUFS, spectral balance, key and BPM — against one of 154 genre presets in 12 families, from trap and drill to gospel and Afrobeats.
Per-stem chains, gain staging, pan, sidechain ducking, vocal rides and mastering, rendered through real VST3 plugins. The mix’s levels, pans, chain order and intensity go live to the Channel Strips in your DAW. Every decision is there to inspect.
Export WAV, MP3, FLAC or AAC, or take the whole session into REAPER, Ableton Live, DAWproject or Pro Tools, built track by track.
New Mix Session — point at your stems or a 2-track, pick from 154 genre presets in 12 families, go.
Short on stems? Press ANALYZE on one track instead — see One ecosystem below.
Hear it
“Guiding Light” was written, performed and produced entirely inside StemDesk: it generated the track, split it into stems, then mixed and mastered them through its own plugins. Flip the switch mid-play and hear exactly what the mix engine did.
Both sides are loudness-matched — what you hear is the mix and the master, not one side simply turned up.
One ecosystem
Put the StemDesk Channel Strip on your tracks and the app does the setup for you. We haven’t seen another mixer do this.

It is a normal plugin. Add it to each track in your DAW, the way you add any plugin.

Name the track, or just play it: the app matches each plugin to its stem and sets the controls the analysis chooses for that track. Gain, trim and the vocal rides live on the strip, so the strip is your fader.

In the app, levels (gain, trim, pan), bypasses and intensity follow into the plugin. Chain order goes one way, from the app to the plugin. Re-run the analysis and every open strip re-configures. You see what every plugin is doing without leaving the app.

Put the Channel Strip on any track, in any DAW, press ANALYZE and play up to 60 seconds. It runs the same per-track analysis the full mix uses and sets up the whole plugin for the role it detects. No bouncing, no session. (Fader rides come from the full-mix flow, which hears every track at once.) Pro tier.
The app runs in the background while you work; the plugins are normal VST3/AU/AAX and work on their own. In Pro Tools you add the Channel Strip to each track yourself — it then configures itself.
The plugins
Open them in Pro Tools, Logic, Live, FL Studio, Cubase, Studio One, Bitwig or REAPER. Three are free. Every interface below is the real plugin, running on audio.
Tap any plugin to see it up close
Six of the seventeen, running on audio in a DAW — real spectrum, real gain reduction, real metering.
Your AI engineer
Twenty-two tools in one panel beside your session, on a local model. It never edits your mix behind your back — every suggestion is shown, explained, and applied only when you say so. Think of it as a second engineer’s opinion: you hear it, then you decide.
A letter grade (A to F): what’s working, what’s masking, issues ranked by severity with the tracks they affect — and one-click “Apply Fixes”, fully undoable.

Per-stem energy across seven bands, conflicts outlined in red, one-click “Fix Masking”. A Track clashes card lists the track pairs that mask each other, with a suggested cut.

Prompt a track in your song’s key and BPM, and StemDesk splits the result into stems — 4 by default, or 6 with guitar and piano — and drops them into your session — actual stems, not a locked stereo file. The lyric writer drafts verses and hooks from one line.
Music generation is an experimental feature (Settings → Labs) that connects to a local ACE-Step install (a free, separate download you point StemDesk at). Everything else on this page works without it.

Session-specific answers as concrete dB moves, ready to apply. Type “kick needs more punch” — a 314-phrase mix vocabulary resolves it instantly; the local model handles the rest.

Also in the panel:
.stemdeskprofile mix profilesSave a mix profile and load it on another session. Vocal AI rides the lead vocal word by word — the speech model ships with StemDesk. The text tools run on a local model you download once, in-app, with your consent; without it the analysis falls back to deterministic rules, so you always get a result, never a spinner.
Inside the mix engine
Nothing is a black box. Your audio stays on your computer unless you choose to connect a remote server.

Describe an idea in one line and the AI drafts verses and hooks. Your lyrics are saved in a history.

Prompt a track in your song’s key and BPM. StemDesk splits the result into stems: 4 by default, or 6 with guitar and piano. Experimental (Settings → Labs); needs ACE-Step, a free separate install.

Split a stereo song into 4 or 6 stems on your own computer, or start from your stems folder, a Pro Tools session or a reference track.

Finds the key, tempo, chords and beat grid of your session, so anything you add fits.
Mix Critique, the masking heatmap and the Mix Coach are shown in Your AI engineer above.

Each stem gets a role (one of 24), a loudness reading and a spectral balance, so the mix starts from facts. A close call is flagged for review, and if you move a track the chain re-thinks itself.

Pick a style, from trap and drill to gospel and Afrobeats. Change your mind later and the targets update on your next Auto-Mix or bounce.

Per-stem chains, levels, pan, sidechain ducking, vocal rides and mastering, in one click (Studio tier). Every decision can be inspected.

From gating to limiting, plus the ride and duck lanes. Each stage turns on only where it helps. Click a stage to bypass it.

Vocal AI transcribes the lead vocal and rides it word by word, around phrases and breaths. The speech model ships with StemDesk.

Checks every stem for silence, clipping and over-compression, and suggests fixes you can apply.

Arrangement analysis of your session, with a grade and suggestions.

Load a commercial track as a reference and match toward it with a real 64-point filter. Save a mix profile and load it on another session.

Eleven delivery targets, from Spotify to CD, plus a custom loudness and true-peak target. A live meter shows where you are. An optional adaptive master EQ adapts the curve verse by verse and chorus by chorus instead of stamping one curve across the song.

After Auto-Mix a card tells you how loud the master came out and which file to upload.

Compare the raw sum, gain and pan only, the processed mix and the master, loudness-matched so the louder one never just wins.

Write presets for FabFilter’s Pro series if you own it, or use StemDesk’s own Channel Strip. You never have to buy another company’s plugins. A chain order editor lets you reorder. Generate All: one click writes presets for the main plugins.

Take the whole mix into REAPER, Ableton Live, DAWproject or Pro Tools (Studio tier). Over 40 tracks, the app suggests grouping them in your DAW first.

Point it at a folder and StemDesk mixes every song, then reports how consistent the record sounds (Studio tier).

Ctrl+Z undoes gain moves, pan moves, applied AI fixes and reference matching. Every plugin has Undo and Redo at the top.

Arrow keys move a fader 0.1 dB, Shift 1 dB, Page Up and Down 3 dB. The full keyboard list is in the manual.

On a Channel Strip you can switch the Tuner or Vocal Cleanup off. Off means not processed, and no added delay.

The AI runs on a model you download once, with your consent, about 8.4 GB. Pick from 11 models sized to your hardware. Without it, the analysis falls back to rules, so you always get a result.

Open Settings to Repair Plugins if one is missing or out of date (it never deletes one), run the First session guide, or check for updates in your own language. StemDesk speaks 25 languages, including right-to-left Arabic and Hebrew.
DAW support
Any DAW that loads VST3, AAX or Audio Units. Pro Tools and REAPER were tested end to end in the versions below; the rest are expected to work as standard hosts.
| Host | Format | Status |
|---|---|---|
| Pro Tools 2026.4.1 (Windows) | AAX · PACE-signed, no iLok | Tested |
| Pro Tools Intro 2026.4.1 (macOS) | AAX · PACE-signed, no iLok | Tested |
| REAPER 7.63 (Windows) | VST3 | Tested |
| REAPER 7.79 (macOS) | VST3 · Audio Unit | Tested |
| REAPER 7.82 (Linux) | VST3 | Tested |
| FL Studio 2026 (Windows) | VST3 | Loaded in v1.0.29; not re-tested since |
| Cubase 12/13 | VST3 | Expected to work; not yet tested in the host |
| Ableton Live 12 Lite 12.3.6 (Windows) | VST3 | Tested |
| Ableton Live 11 | VST3 | Expected to work; not yet tested in the host |
| Studio One 6 | VST3 | Expected to work; not yet tested in the host |
| Bitwig Studio 5 | VST3 | Expected to work; not yet tested in the host |
| Logic Pro (macOS) | AU | Passes Apple’s auval; not yet tested in Logic |
| GarageBand 10.4.14 (macOS) | AU | Tested |
Studio tier: REAPER, Ableton Live, DAWproject, Pro Tools.
Studio tier builds your session for you:
| Target | What you get |
|---|---|
| REAPER (.rpp) | Full session — tracks, levels, pans, buses, and real plugin instances preloaded with the mix's computed settings (StemDesk plugins, or your own FabFilter) |
| Pro Tools (live) | StemDesk drives a running Pro Tools and builds the session track-by-track |
| Ableton Live (.als) | Set with your stems, song tempo, and either a FabFilter chain or a StemDesk Channel Strip on every track |
| DAWproject | Open interchange format for Studio One and other supporting hosts |
Every export comes two ways: around FabFilter's Pro series if you own it, or around StemDesk's own Channel Strip if you don't — you never have to buy another company's plugins to get a finished, editable session. Over 40 tracks, the app suggests grouping them in your DAW first.
How StemDesk compares
Plenty of tools auto-master a stereo file. A few auto-mix stems in the cloud. StemDesk mixes your stems on your own machine, drives the plugins in your DAW, and hands you back the session.
| Capability | StemDesk | RoEx Automix | Cryo Mix | iZotope Neutron 5 | iZotope Ozone | LANDR |
|---|---|---|---|---|---|---|
| Full mix from your multitrack stems | ✓ native — a Channel Strip chain per stem | ✓ up to 32 stems | ✓ | ✗ per-track assistant inside your DAW | ✗ masters a stereo file | ✗ masters a stereo file |
| Hands back an editable DAW session | ✓ Studio tier: REAPER and Ableton sessions open with the plug-ins already set; DAWproject carries the plug-in settings (not yet tested in Bitwig or Studio One); Pro Tools built live from processed stems | partial — processed stems and DAW exports | ✗ browser only | n/a — lives inside your session | n/a — lives inside your session | — |
| Audio stays on your machine | ✓ by default — nothing uploads | partial — cloud on the web app; a Desktop app is included with Pro | ✗ runs in the browser | — | — | ✗ cloud upload |
| Included plugin suite, usable in any VST3, AAX or AU DAW | ✓ 17 plugins on Pro and Studio (Compressor, Gate and Limiter on Free) | — | ✗ | partial — 10 modules in one plugin | partial — mastering modules | partial — 70+ plugin bundle |
| AI critique, coaching & spectral analysis | ✓ full local panel — critique, coach, heatmap, diagnostics | partial — mix reports | partial — chat assistant | partial — Mix Assistant | partial — Master Assistant | — |
| Works fully offline | ✓ | partial — Desktop app | ✗ | — | — | — |
| Batch album processing | ✓ Studio tier | — | — | — | — | — |
| Own it outright — no subscription | ✓ one-time license, $0 / $99 / $179 | ✗ subscription | ✗ subscription | ✓ one-time purchase option | ✓ one-time purchase option | ✗ subscription |
As of v1.0.53, compiled from each vendor's own published pages. Products, editions, and pricing change often — check with the vendor before relying on any cell here. "—" means we could not confirm a published capability as of that version, not that it is absent. Every product above is good at what it is built for; this table is about what each one is built for, not which is better made. All product names and trademarks are the property of their respective owners; no affiliation, sponsorship, or endorsement is implied.
Pricing
One-time license, activated offline. No subscription, no iLok — ever. Start free, upgrade when you want the full suite. Every v1.x update is free, in writing (EULA §3).
Try StemDesk free before you buy. Paid licenses are refunded for duplicate purchases or technical problems we can’t fix, not for a change of mind. Your legal rights still apply, including the EU/UK 14-day cancellation right. Refund policy →
The creator
Twenty years of a working studio’s method, compiled into software.
Built from nothing into a premier Orange County studio, est. 2005, now in the Inland Empire — SSL Sigma and Pro Tools Ultimate HDX. Twenty-plus years of records: Grammy winners, Oscar-nominated actors, platinum artists, Netflix, Gold Records, JBL/NBA, Power Rangers, audiobooks for Penguin Random House and Simon & Schuster, and voice-over for Blizzard and the U.S. FTC.
Also a computer scientist and a licensed California attorney. Built and maintained by Asaf Fulks, a working studio engineer since 2005, using AI-assisted development and open-source components.
The plugins were voiced against what the SSL room at OC Recording does.
The separation pipeline, the mix engine and all seventeen plugins were designed by the person who mixes on them.
An 858-page textbook and a certificate course — the reasoning behind every move is taught, not hidden behind “AI”.
“Your audio stays on your machine unless you send it somewhere” was written by someone professionally obligated to mean it.
FAQ
No. Analyze Track configures a Channel Strip from up to 60 seconds of playback, with the app running in the background — press ANALYZE, play, done. Bouncing is only needed for the full-mix flow: whole-song analysis, fader rides and a rendered mix. Both are walked through step by step in the user manual.
Yes. With the app running, each Channel Strip is matched to its track and gets its settings live; faders and bypasses follow the app, and meters stream back. In Pro Tools you add the Channel Strip to each track yourself.
Yes, they are normal VST3/AU/AAX plugins. The AI button and the live settings need the app running.
The plugins load in any DAW that hosts VST3, AU or AAX — Pro Tools, Logic, Ableton Live, FL Studio, Cubase, Studio One, Bitwig, REAPER, GarageBand. Session export (Studio tier) goes to REAPER, Ableton Live, DAWproject (Studio One, Bitwig and others) and Pro Tools, built live. The host table shows what was tested end to end.
No. Separation, analysis, mixing and the AI assistant run locally by default, and nothing is uploaded. If you choose to connect your own remote AI server, prompts, mix measurements and track names (never audio) go to it; music generation on a remote ACE-Step server sends it your prompt, and Cover sends it the reference song. AI models and weights download once, on first use.
For the download, the optional AI model, and a one-time fetch of extra model files the first time you use separation or vocal restoration (and, on a Mac, the first-run components). After that, mixing, separation and the plugins work fully offline.
The mix itself is built by deterministic, measurable DSP. The local AI suite sits on top — graded critique, a session-aware coach, per-stem diagnostics, vocal transcription, reference matching, preset generation, and music & lyric generation — all optional, all user-driven, never a black box in your render path.
Settings → Repair Plugins copies the right version back. It never deletes a plugin. Upgrades also replace older plugins automatically, and the installer asks you to close your DAW first if it has a StemDesk plugin open.
Yes. A plain-English first-session guide is built in: Settings → First session guide. The full keyboard shortcut list is in the user manual.
No. StemDesk plug-ins never need an iLok. The AAX versions are PACE-signed, as Avid requires, but not iLok-licensed, so no iLok account, dongle or cloud authorization is needed for StemDesk in Pro Tools or any other DAW. (Pro Tools itself may ask for an iLok account for Avid’s own licensing.)
A key arrives by email after purchase and activates offline inside the app — cryptographically signed, no account, no phone-home.
Try StemDesk free before you buy. Paid licenses are refunded for duplicate purchases or technical problems we can’t fix, not for a change of mind. Your legal rights still apply, including the EU/UK 14-day cancellation right. Refund policy →
Purchases made before 2 November 2026 keep the refund terms in force when you bought (14-day refund for any reason).
Until our checkout moves to Stripe Managed Payments, purchases are processed by Fulks, Inc. through Stripe, and the references to Link as merchant of record in the Terms and EULA apply from that move.
Music generation is an experimental feature, off by default: switch it on in Settings → Labs. It runs through ACE-Step, a free local generator you install separately and point StemDesk at — we don't bundle it, so the installer stays lean for the majority who won't use it. Everything else, including the lyric writer, critique and coach, uses the model StemDesk downloads in-app.
The default local model runs on modest hardware. Power users can pick from a catalog of 11 local models (2.8–30.4 GB), each recommended against your detected VRAM — and a multi-GPU machine is sized against its combined VRAM, not one card.
Yes — macOS on Apple Silicon (M1 or later) is in beta, signed and notarized, with all 17 plugins in both VST3 and Audio Unit formats. The AUs pass Apple’s auval and are Developer ID signed, notarized and stapled. All 17 load and play in GarageBand; Logic Pro has not been tested yet. External sidechain in AU requires a stereo main input. AAX for Pro Tools on Windows and macOS.
Latest: v1.0.53 · November 2026
Try StemDesk free before you buy. Paid licenses are refunded for duplicate purchases or technical problems we can’t fix, not for a change of mind. Your legal rights still apply, including the EU/UK 14-day cancellation right. Refund policy →
One installer for every tier. Your license key unlocks Pro or Studio inside the app. New here? Read the user manual and the release notes.
Windows 10 (version 1809, October 2018 Update, or later) / Windows 11, 64-bit · v1.0.53. The installer and plugins are code-signed; verify with the SHA-256 checksum. AAX plug-ins are PACE-signed for Pro Tools and need no iLok account or authorization. The optional local AI model is a one-time ~8.4 GB download, fetched only with your consent. The app, plug-ins and installer speak 25 languages — see the release notes.
Arabic and Hebrew use a right-to-left layout. Meters, faders and waveforms keep their normal direction.
Upgrades replace older plugins automatically; the installer asks you to close your DAW first if it has a StemDesk plugin open. If an installed plugin is missing or out of date, Settings → Repair Plugins copies the right version back — it never deletes a plugin. Settings → Check for Updates tells you about a new version in your language. A plain-English first-session guide is built in: Settings → First session guide.
Windows SmartScreen may show “Windows protected your PC” — normal for new software while Microsoft builds download reputation for it. Click More info, confirm the publisher reads “Fulks, Inc.” (that’s us), then Run anyway.
On Windows the installer adds a small helper, the Plug-in Manager, that keeps your plug-in folders in step with your license. It runs with administrator rights as a Windows Scheduled Task that the installer registers, and it copies or removes only StemDesk's own VST3 and AAX plug-ins: the three free ones (Compressor, Gate and Limiter) on the Free tier, all 17 on Pro and Studio. Studio adds app features, not extra plug-ins. It never touches other makers' plug-ins. On macOS and Linux there is no helper: the app copies the plug-ins itself.
macOS (beta): StemDesk-1.0.53-arm64.dmg — Apple Silicon (M1 or later), macOS 26.6+. Signed and notarized by Apple, so it opens without warnings: drag StemDesk to Applications and launch. The 17 plugins install as VST3 to ~/Library/Audio/Plug-Ins/VST3 and as Audio Units to ~/Library/Audio/Plug-Ins/Components; the AAX versions for Pro Tools install to /Library/Application Support/Avid/Audio/Plug-Ins/StemDesk. Same license keys; the optional AI model downloads in-app with your consent. First run on a Mac also downloads the audio-processing components (over a gigabyte) from PyPI, so an internet connection is required once. Verify: SHA-256.
Support for earlier macOS versions is planned for the next update.
Linux (beta): StemDesk-1.0.53.AppImage — one file, everything included. x86_64, glibc 2.28+ (Ubuntu 22.04+/Debian 12+ and similar). Runs on a stock Ubuntu Desktop install — v1.0.24 and later work with the FUSE 3 stack your distro already ships (releases before v1.0.24 needed a one-time sudo apt install libfuse2t64 on Ubuntu 24.04+). On Ubuntu 24.04 and later the Chromium sandbox is disabled for the AppImage, an AppImage limitation; StemDesk renders only its own bundled interface. chmod +x, run, and the 17 plugins install to ~/.vst3 as VST3 (AAX (Pro Tools) is available on Windows and macOS). Same license keys work; the optional AI model downloads in-app with your consent, exactly as on Windows. Verify: SHA-256 · GPG signature (release key, fingerprint 0213 8837 D626 85A4 0CA4 1265 D36C 257A C65B A130).
An AppImage has no uninstaller. Delete the AppImage file, then delete whichever of these you want gone.
~/.config/stemdesk — Settings, your licence and working caches; also the AI music-generation environment if you set that up. (1 GB or more; app files)~/.local/share/StemDesk — The AI runtime (the engine that runs the language model) and its working files. (1 GB or more; app files)~/.local/share/stemdesk — The plug-in manager's log files. (under 100 MB; app files)~/Documents/StemDesk — Your sessions, mixes, presets and profiles. These are yours: keep them, or delete the folder once you have saved anything you want. (1 GB or more; your files)<each session folder>/Presets/.superseded — Old preset files the app set aside. This is your data; it is safe to delete. StemDesk keeps the last 3 runs there and removes older ones itself. (under 100 MB; your files)~/Documents/StemDesk/Models — Downloaded models (the stem-separation model and the speech models); they download again if needed. This folder is inside the one above, so you can delete just it and keep your sessions. (100 MB to 1 GB; downloaded model)~/.cache/stemdesk — Temporary working files; StemDesk deletes the ones older than 7 days by itself. If that folder cannot be made private to you, StemDesk uses the system temporary folder (/tmp) instead, and the system clears it. (100 MB to 1 GB; app files)~/.cache/StemDesk — A small cache of plug-in file fingerprints, so StemDesk can quickly tell whether your plug-ins changed. It is rebuilt by itself if you delete it. (under 100 MB; app files)~/.cache/voicefixer — The voice-repair model, downloaded the first time you use it. (100 MB to 1 GB; downloaded model)~/.cache/torch — A model cache, only if it exists. (100 MB to 1 GB; downloaded model)~/.cache/huggingface — StemDesk versions before 1.0.53 kept the stem-separation model here. Other AI tools share this folder, so delete it only if you use none of them. (100 MB to 1 GB; downloaded model)~/.vst3/StemDesk — The StemDesk VST3 plug-ins. (100 MB to 1 GB; app files)The language models are in Ollama's shared store (~/.ollama/models), so do not delete that folder: other apps may use it. Remove StemDesk's models by name instead. They all start with stemdesk-; the default is stemdesk-gemma4-e4b-q4, and any larger model you chose in Settings is a separate entry that can be tens of GB.
Start the Ollama that StemDesk installed, and leave it running in a terminal: "$HOME/.local/share/StemDesk/runtime/ollama/ollama" serve
In a second terminal, list the models: "$HOME/.local/share/StemDesk/runtime/ollama/ollama" list
Remove each model that starts with stemdesk-, for example: "$HOME/.local/share/StemDesk/runtime/ollama/ollama" rm stemdesk-gemma4-e4b-q4
StemDesk never deletes ~/Documents/StemDesk for you: it holds your sessions and mixes, so it stays until you delete it yourself.
There is no uninstaller. Drag StemDesk from Applications to the Trash, then delete whichever of these you want gone.
~/Library/Application Support/stemdesk — Settings, your licence, the Python libraries, the AI runtime and working caches; also the AI music-generation environment if you set that up. (1 GB or more; app files)~/Documents/StemDesk — Your sessions, mixes, presets and profiles. These are yours: keep them, or delete the folder once you have saved anything you want. (1 GB or more; your files)<each session folder>/Presets/.superseded — Old preset files the app set aside. This is your data; it is safe to delete. StemDesk keeps the last 3 runs there and removes older ones itself. (under 100 MB; your files)~/Documents/StemDesk/Models — Downloaded models (the stem-separation model and the speech models); they download again if needed. This folder is inside the one above, so you can delete just it and keep your sessions. (100 MB to 1 GB; downloaded model)~/.cache/StemDesk — A small cache of plug-in file fingerprints, so StemDesk can quickly tell whether your plug-ins changed. It is rebuilt by itself if you delete it. (under 100 MB; app files)~/.cache/voicefixer — The voice-repair model, downloaded the first time you use it. (100 MB to 1 GB; downloaded model)~/.cache/torch — A model cache, only if it exists. (100 MB to 1 GB; downloaded model)~/.cache/huggingface — StemDesk versions before 1.0.53 kept the stem-separation model here. Other AI tools share this folder, so delete it only if you use none of them. (100 MB to 1 GB; downloaded model)~/Library/Audio/Plug-Ins/VST3/StemDesk — The StemDesk VST3 plug-ins. (100 MB to 1 GB; app files)~/Library/Audio/Plug-Ins/Components — The StemDesk Audio Unit plug-ins (the files named StemDesk ... .component). Other makers' plug-ins share this folder, so delete only ours. (100 MB to 1 GB; app files)/Library/Application Support/Avid/Audio/Plug-Ins/StemDesk — The StemDesk Pro Tools (AAX) plug-ins, installed only if you have Pro Tools. (100 MB to 1 GB; app files)~/Library/Caches/StemDesk Channel Strip — Temporary files from the Channel Strip plug-in's track analysis; emptied and removed when the plug-in closes. (under 100 MB; app files)~/Library/Audio/Presets/StemDesk — Empty preset folders that your host creates for the Audio Unit plug-ins; safe to delete. (under 100 MB; app files)~/Library/Preferences/com.stemdesk.app.plist — The app's preferences file. (under 100 MB; app files)To remove the Pro Tools plug-ins, run this in Terminal (it asks for your administrator password): sudo rm -rf "/Library/Application Support/Avid/Audio/Plug-Ins/StemDesk"
The language models are in Ollama's shared store (~/.ollama/models), so do not delete that folder: other apps may use it. Remove StemDesk's models by name instead. They all start with stemdesk-; the default is stemdesk-gemma4-e4b-q4, and any larger model you chose in Settings is a separate entry that can be tens of GB.
Start the Ollama that StemDesk installed, and leave it running in a terminal: "$HOME/Library/Application Support/stemdesk/runtime/ollama/ollama" serve
In a second terminal, list the models: "$HOME/Library/Application Support/stemdesk/runtime/ollama/ollama" list
Remove each model that starts with stemdesk-, for example: "$HOME/Library/Application Support/stemdesk/runtime/ollama/ollama" rm stemdesk-gemma4-e4b-q4
StemDesk never deletes ~/Documents/StemDesk for you: it holds your sessions and mixes, so it stays until you delete it yourself.
Every AI feature in StemDesk runs on your machine by default — your audio stays on it unless you point StemDesk at a remote server. The AI assistant, mix critique and arrangement features run on your choice of open-weight language model, downloaded in-app with your consent, checksum-verified, and matched to your hardware automatically: Google Gemma 4 — E2B · E4B (the default) · 26B MoE · 31B (up to near-lossless Q8 for multi-GPU rigs) • Alibaba Qwen 3.5 — 4B · 9B · 35B MoE • Qwen 3.6 — 27B • Qwen 3.8 — 27B, supported the week of its release, up to near-lossless Q8 for multi-GPU rigs.
Also on board: Demucs (Meta’s stem-separation model) splits full mixes into stems, and the optional ACE-Step setup (one click on Windows with an NVIDIA GPU) adds AI music generation that runs on your machine. All open-weight and all local after a one-time download — no account, and no cloud unless you add one.