StemDesk User Manual
For StemDesk v1.0.53 on Windows, macOS and Linux. This is the authoritative description of how the product works in this version.
- The two halves of StemDesk
- Install and first run
- Licenses and tiers
- The plug-ins
- The Channel Strip
- Workflow A — Analyze Track (no bouncing)
- Workflow B — the full mix, part 1: set up a session
- Workflow B, part 2: the results screen and Auto-Mix
- Exporting audio and DAW sessions
- The AI Intelligence panel
- Settings
- Keyboard shortcuts
- Platform notes and system requirements
- Troubleshooting
- Getting help
1. The two halves of StemDesk
The plug-ins are ordinary DAW plug-ins. You put them on tracks and they process audio live, like any plug-in you own. There are 17 of them (section 4).
The desktop app is the mixing brain. It separates a 2-track into stems, analyzes audio, decides what every track needs, renders full mixes, and drives your plug-ins directly: every StemDesk plug-in you have open connects to the app over a private connection on your own machine (127.0.0.1, port 9800) and receives its settings from the app.
There are two ways to get from a track to a configured mix:
- Workflow A — Analyze Track. Put the Channel Strip on a track, press ANALYZE, play up to 60 seconds. The app configures the whole strip. No bouncing, no session. Sections 5 and 6.
- Workflow B — the full mix. Bounce stems (or split a 2-track), build a session in the app, run Auto-Mix, and hand the result back to your DAW. Sections 7 to 9.
2. Install and first run
Installing
| Platform | Installer | Plug-in formats you get | Where the plug-ins go |
|---|---|---|---|
| Windows 10 (1809+) / 11, 64-bit | StemDesk-Setup-v1.0.53.exe. Installs for all users and asks for administrator approval. SmartScreen may show "Unknown publisher" the first time; the signer is Fulks, Inc.. Click More info → Run anyway. | VST3 and AAX | C:\Program Files\Common Files\VST3\StemDesk and ...\Avid\Audio\Plug-Ins\StemDesk |
| macOS 26.6+, Apple Silicon (M1 or newer) | StemDesk-1.0.53-arm64.dmg, signed and notarized by Apple. Drag StemDesk to Applications. | VST3 and Audio Unit | ~/Library/Audio/Plug-Ins/VST3 and ~/Library/Audio/Plug-Ins/Components |
| Linux, x86-64 | StemDesk-1.0.53.AppImage, GPG-signed. Make it executable and run it. | VST3 | ~/.vst3/StemDesk |
What the installer does, and what the app does
The installer puts the three free plug-ins (Compressor, Gate, Limiter) in your plug-in folders. The other 14 plug-ins are installed by the app the first time it starts with a verified Pro or Studio license. So the order that works is: install → start the app → activate your license (section 3) → then restart your DAW or rescan plug-ins. If you restart the DAW before activating, it will see only the free trio.
Your first launch, in order
- License agreement. A full-screen page with EULA, Terms and Privacy tabs. You must scroll each to the bottom before Accept is enabled. Decline closes the app.
- "Set up StemDesk's AI". StemDesk needs one download before its AI features work, and it runs entirely on your computer: the AI model, a one-time transfer of about 8.4 GB from StemDesk's content delivery network. StemDesk picks the version that suits your graphics card (section 13). It powers Ask AI, the mix critique, Auto-Mix suggestions and lyric writing. No download is needed to mix: Auto-Mix works on your own stems right away. Choose Download now or Not now. With Not now, everything except the AI features still works; you can start the download later from Settings → LLM Model. Until it finishes, the AI features do nothing. That is expected, not a fault. Offline or air-gapped machines can import the model by hand; see the sideload guide linked from that dialog. Music generation (ACE-Step) is an experimental feature you switch on in Settings → Labs; it is not part of this prompt.
- Component check. The splash screen verifies the bundled Python components ("Verifying components…"). On macOS the app installs its Python dependencies itself on first run, which takes a few minutes.
- Onboarding tour. Six short steps: session setup, mix targets, Auto-Mix, playback, what Free, Pro and Studio include, sessions and profiles. Skip it any time.
Two more downloads happen later, on demand: stem-separation weights (Demucs) the first time you split a track, and, if you turn on the Labs music generator, ACE-Step quality tiers the first time you pick them.
3. Licenses and tiers
| Free | Pro | Studio | |
|---|---|---|---|
| Plug-ins | Compressor, Gate, Limiter | All 17 | All 17 |
| Desktop app core: stem separation from a 2-track, analysis, mix targets, 154 genre presets in 12 families, the local AI suite, WAV / MP3 / FLAC / AAC export | ✓ | ✓ | ✓ |
| Analyze Track (needs the Channel Strip) | — | ✓ | ✓ |
| Auto-Mix (the rendered full mix) | — | — | ✓ |
| Session export to REAPER, Ableton Live, DAWproject; Pro Tools session build | — | — | ✓ |
| Batch album processing | — | — | ✓ |
Studio does not add plug-ins; its difference is the app's rendering and delivery features. Every locked feature shows a 🔒 badge and opens an upgrade dialog instead of failing silently. Pricing is on stemdesk.ai; every v1.x update is free.
Activating
Your license key arrives by email after purchase. In the app, click the StemDesk logo at the top-left (or press Ctrl+Shift+L, Cmd on a Mac), paste the key and press Activate (Ctrl+Enter also works). The same dialog is reachable from Settings → License → Manage License. The tier shows as a badge in the title bar.
Moving to another computer: open the same dialog and press Deactivate to remove the license from this machine first.
4. The plug-ins
| Plug-in | Tier | What it is |
|---|---|---|
| Compressor | Free | Compressor with sidechain high-pass / low-pass filters, lookahead, hold, range, stereo link and a parallel Mix control. |
| Gate | Free | Gate with the same sidechain filtering, lookahead, hold and range. |
| Limiter | Free | True-peak limiter with ceiling, lookahead and release. |
| EQ | Pro | 8-band parametric with dynamic EQ per band, Mid/Side and Left/Right placement, auto-gain, a clickable piano strip and a spectrogram view. |
| Multiband Comp | Pro | Multiband compressor. |
| De-Esser | Pro | Sibilance control with a DET meter for the detection band and Listen buttons for the detection band and for what is being removed. |
| De-Noise | Pro | Broadband noise reduction. |
| Saturator | Pro | Eight saturation modes (Tube, Tape, Clean, Transformer, Fuzz, Diode, Amp, Crush). Drive at exactly 0 adds no saturation. |
| Reverb | Pro | Room, plate, hall and ambient algorithms. |
| Delay | Pro | Tempo-synced, filtered delay with ducking. |
| Fattener | Pro | Chorus, micro-shift and Haas widening. |
| Transient Shaper | Pro | Attack and sustain shaping. |
| Stereo Imager | Pro | Per-band stereo width. |
| Drum Tuner | Pro | Granular pitch shifting for drums. |
| Vocal Cleanup | Pro | Breath, click and noise cleanup. |
| Vocal Tuner | Pro | Pitch correction: Engage, Key and Scale (Chromatic, Major, Minor), Retune in ms (0 is a hard snap, higher glides), Amount, and factory presets from Hard Tune to Natural Touch-Up. |
| Channel Strip | Pro | All of the above in one reorderable 13-module strip, plus ANALYZE, AUTO and Ask AI. Section 5. |
Tested: Pro Tools 2026.4.1 and REAPER 7.63 on Windows; Pro Tools Intro 2026.4.1 and REAPER 7.79 on macOS; REAPER 7.82 on Linux. Cubase 12 and 13, Ableton Live 11 and 12, Studio One 6, Bitwig 5 and FL Studio 2026 are expected to work as standard VST3 hosts (FL Studio loaded the free plug-ins in v1.0.29 and has not been re-tested since). On macOS the AU versions pass Apple's validation (auval), and all 17 load and play in GarageBand; Logic Pro has not been tested yet. The in-app compatibility matrix (Settings → View DAW Compatibility) lists which hosts are tested and which are expected.
Oversampling. Saturation creates harmonics, and any harmonic above half the internal sample rate folds back into the audible range as aliasing. The Saturator can run at 1x, 2x, 4x (the default) or 8x its sample rate: 8x is the cleanest when driven hard and uses the most CPU; 2x and 1x use less CPU and alias more. Crush is a sample-rate reducer, so aliasing is what it does.
The table shows the strongest alias line in dB below the signal (the lower, the cleaner) for sine tones from 3 to 12 kHz at 44.1 and 48 kHz. Typical is Drive 25 with Mix 25; extreme is Drive 100 with Mix 100.
| Mode | 4x, typical | 8x, typical | 4x, extreme | 8x, extreme |
|---|---|---|---|---|
| Tube | -83 | -82 | -44 | -63 |
| Tape | -92 | -93 | -64 | -73 |
| Clean | -85 | -84 | -47 | -66 |
| Transformer | -71 | -78 | -30 | -59 |
| Amp | -73 | -80 | -40 | -53 |
| Fuzz | -51 | -63 | -35 | -42 |
| Diode | -57 | -68 | -40 | -49 |
| Crush | by design | by design | by design | by design |
Tape's extra alias reduction is on at 2x, 4x and 8x, and at 1x from 88.2 kHz up. At 1x with 44.1 or 48 kHz, Tape keeps its earlier sound (the alias reduction would dull its top octave there), so choose 2x or more for the cleanest Tape.
Decay Hi multiplies each mode's own high-frequency decay. At high Damping the longest setting is limited by the Damping itself (at Damping 50, x4 reaches about 1.4x to 2.8x depending on the mode; at Damping 0, 2.6x to 4x).
5. The Channel Strip
The strip is 13 processors in one plug-in: Vocal Cleanup, Gate, EQ, Compressor, De-Esser, Saturator, Fattener, Transient Shaper, Stereo Imager, Drum Tuner, Delay, Reverb and Limiter. That is also the default order; the app changes the order per detected role, and you can drag a module button in the left sidebar to reorder it yourself.
The header buttons
| Button | Needs the app? | What it does |
|---|---|---|
| ANALYZE | Yes | Captures this track's audio, sends it to the app, and the app configures every module and the chain order for the detected role. Section 6. |
| AUTO | No | Local auto-configure. The strip reads its own recent level, peak and crest measurements (or uses conservative defaults if no signal has played) and sets every processor itself. Faster and works offline, but it is not the app's analysis: no role detection, no chain order for the role. |
| AI | Yes, plus the AI model download | Opens the chat bar. Type a request such as "make the vocal sit further back", press Send, and the app adjusts the strip's processors and reports what it changed. |
Also in the header: the connection dot (green when linked to the app), the track-name field, preset previous / select / next / save, and a skin toggle. Undo and Redo live in the strip's bottom bar. Everything the app changes is written as real parameter values, so your DAW's own undo and automation see them too.
6. Workflow A — Analyze Track (no bouncing)
- Start the StemDesk app and leave it running. Put a Channel Strip on the track (Pro license or higher).
- Press ANALYZE. The button turns orange and starts counting:
12/60s. - Play the track through the strip. Up to 60 seconds are captured; a verse plus the loudest chorus is ideal. You can stop early; anything over one second is analyzed.
- When the capture ends, the button reads ANALYZING while the app runs the same per-track analysis the full mix uses.
- The button then reports CONFIGURED for a few seconds. Hover it for the detail (for example "Configured: lead vocal (vocal)").
What you get
Input trim and gain staging, the 13-module chain order for the detected role (cleanup first, then dynamics, EQ and tone in the order that role wants), EQ bands, compressor, de-esser, gate, saturation, limiter and the rest, applied to the strip as parameter settings you can see and undo in your DAW.
Honest limits
- It analyzes the up-to-60 seconds you played, not the whole song. Treat it as a fast, smart per-track setup pass and re-run it on a better section if the result is off.
- It does not write fader rides or duck automation. Rides need whole-mix context, which one track cannot give. Use Workflow B.
- It does not rename the strip; track naming happens in Workflow B when the app assigns stems to strips.
Messages you may see on the button
| Label | Meaning (full text in the button's tooltip) |
|---|---|
| APP OFFLINE | The app is not running or not reachable on port 9800. Start it, or use Settings → Channel Strip Server → Restart. |
| NO AUDIO | Less than one second of audio reached the strip during the capture. Start playback, then press ANALYZE again. |
| FAILED | The app refused or could not analyze the capture. The most common cause is a capture that is genuinely silent. The tooltip carries the reason. |
| NO APP REPLY | The app did not answer within two minutes. Check that it is still running, then try again. |
| REINSTALL | The app rejected this plug-in because the two are from different versions. Reinstall StemDesk so the plug-ins match the app. |
The capture is a temporary file in your system temp folder (about 23 MB per minute). The app deletes it after analysis.
7. Workflow B — the full mix, part 1: set up a session
Press + New Session in the left sidebar. The setup screen has three parts.
Source
- Stems Folder (required): a folder of WAV files, one per track or stem group, bounced clean with no master-bus processing.
Name your stems. StemDesk uses the file names ("Kick", "808", "Vox", "Snare") to classify roles and set mix targets. Generic names like "Audio 1" or "Track 3" force a fallback to spectral analysis, which is less accurate. Rename before you run. The file name also becomes the track name the app assigns to each strip. If no track is recognized as bass, no track gets the bass-specific settings (the level target, EQ and compression for the low end); name your bass track "Bass", or set its role to Bass in the track list.Big sessions. StemDesk works best with about 8 to 40 tracks. For a project with many more (a 100-track session full of synth layers, one-shots and vocal doubles), first combine tracks into groups in your DAW (drums, bass, guitars, keys, synths, lead vocal, backing vocals, effects) and export one stem per group. Every part is still there, and each stem gets a clear role. Above 40 tracks the app shows this advice above the track list.
- Pro Tools Session (optional): a
.ptxor.ptsfile whose track metadata the pipeline uses. - Reference Track (optional): a commercial reference for the reference-matching tools in the AI panel (section 10).
- Split a 2-Track: choose a stereo file and 4 or 6 stems. Separation runs locally (the first run downloads the separation weights), shows elapsed time and progress, can be cancelled, and writes the stems to a
Stemsfolder next to the input file. - Batch Album Studio: process a whole album of songs with one genre and reference, with a progress view and an album report.
- ✦ Create with AI instead: jumps to music generation (section 10, shown when Settings → Labs → Make a song is on) instead of loading stems.
Pipeline steps
Check boxes for Analysis, Gain, Pan, Rides (volume-ride automation), Vocal AI (word-level vocal alignment with WhisperX), Sidechain Duck (ducks non-kick stems at kick transients) and the AI Modules: Critique, Diagnostics, Arrangement, Full Report and Spectral.
Session name and genre
Name the session, then pick one of the 154 genre presets in 12 families from the searchable chip list. The preset sets the target curve and balance the analysis mixes toward, and the mix engine tunes per family. Press ▶ Start Mix Session. While it runs you can Cancel or show the raw log.
8. Workflow B, part 2: the results screen and Auto-Mix
Mix targets
Every stem gets a row: gain, pan, role, group, intensity override, solo, mute, and an exclude checkbox. Filter rows with Ctrl+F, select all with Ctrl+Shift+A, undo and redo with Ctrl+Z / Ctrl+Y. Excluded stems can be restored. Re-analyze (the ↻ icon in the title bar) re-runs the analysis on the open session.
Changing a track's role re-plans its gain, pan and processing. The processor dots on each row (gate, EQ, compression, saturation, de-esser, delay, reverb, transient, imager, drum tuner and limiter) apply the same way in Auto-Mix and in the Channel Strip in your DAW; click a dot to bypass that processor in both.
The options row
- Intensity: how hard the AI processes, global default 30%, adjustable with ± and overridable per track (0 / 25 / 50 / 75 / 100%). This is the first control to reach for when a mix is too much or not enough. The same intensity reaches the Channel Strips in your DAW.
- Bypass Master: skips the master-bus EQ, glue compressor, multiband and limiter.
- Adaptive EQ: the master EQ follows song sections (verse, chorus). Render-only: it is never written into exported sessions.
- Rides, Vocal AI, Sidechain Duck: the same toggles as the pipeline.
- Mastering target: None, Spotify −14, Apple −16, YouTube −14, Tidal −14, SoundCloud −14, Amazon −14, Deezer −15, Club −8, CD −9, Broadcast −23, Podcast −16 (all LUFS), or Custom with your own LUFS and true-peak.
Driving your plug-ins
With the Channel Strip Server running (it starts with the app; Settings → Channel Strip Server has Start / Stop / Restart), the app matches each open strip to its stem by audio, names the strip after the stem file, pushes the full configuration, and, when Rides is on, writes volume-ride and duck automation to the strip.
Auto-Mix Studio
Auto-Mix renders the stereo mix through real VST3 plug-ins: per-stem chains, gain staging, pan, ducking, rides and the mastering target you chose. Which plug-ins do the rendering depends on Settings → Plugin Set: Auto uses FabFilter's Pro series if it is installed and the StemDesk plug-ins otherwise; you can force either. Regrade re-grades an existing render without re-mixing.
Listening and judging
- The transport plays the multitrack session with meters and a master volume. Space is play / pause; S and M solo and mute the track under the mouse.
- A/B (Tab) flips between raw and mixed; Match loudness-matches the two so you judge balance, not volume.
- Show Auto-Mix opens a compare panel with four sources: Pure Raw (a straight sum), Raw (gain and pan only), Mix (fully processed) and Master (mastered), with its own loudness match.
- Accept All / Reject All applies or discards the AI's per-track suggestions; a feedback box lets you tell the AI what to change.
Mix profiles
Save the current state as a named Mix Profile (save, rename, delete, new) to compare versions inside a session. A profile can be exported as a portable file and imported into another session from the AI panel's Profiles tab.
Sessions save automatically under Documents\StemDesk (relocatable in Settings → Session Storage); Ctrl+S saves on demand. The 📂 icon opens the session, stems or Pro Tools folder.
9. Exporting audio and DAW sessions
The ▾ menu next to Auto-Mix.
Audio
- Export Stems to Folder copies the session's stems.
- Export Mix and Export Master as WAV, MP3, FLAC or AAC (.m4a), with a bitrate choice for MP3 and AAC. Mix items enable after an Auto-Mix has completed; Master items enable only if that Auto-Mix ran with a mastering target.
Which file to use, and how loud it is
| File | What it is | Use it for |
|---|---|---|
| Master | The finished song, mastered to the Final loudness you chose (Spotify −14 LUFS by default) with peaks held below the true-peak ceiling (−1 dB for Spotify). | Uploading to Spotify, Apple Music, YouTube, SoundCloud and the rest, or playing out. |
| Mix | The same mix with no mastering: quieter and fully dynamic. It is never clipped. | Sending to a mastering engineer. |
You do not set the output level by hand. StemDesk masters to the target in the Final loudness menu and, after Auto-Mix, shows the Master's measured loudness and peak and the Mix's peak. Spotify −14 LUFS suits every major streaming service; pick a different target only if you know you need it (Club −8, CD −9, Broadcast −23, or Custom).
"Keep it under −6 dB" or "−9 dB" is advice about headroom: the room left below 0 dB in a mix you hand to a mastering engineer, so they have space to work. It applies to the Mix file, not to a finished Master. Many engineers ask for peaks around −6 to −3 dB; if your Mix peaks higher, they can simply turn it down, because it is not clipped.
DAW sessions Studio
Each target comes in two flavours, FabFilter chain or SD Channel Strip, matching Settings → Plugin Set. A "Before you export" dialog states what survives; this is what it says:
| Target | What carries over | What does not |
|---|---|---|
| REAPER (.rpp) | Track fader and pan; group folders. In FabFilter mode the FabFilter preset state is embedded, so the session plays the AI mix without StemDesk running. | Bus-level fader/pan and group sends; the mastering chain. |
| Ableton Live (.als) | Track fader and pan; plug-ins inserted. | Plug-ins are inserted at their default state (no preset state embedded); no groups, buses or sends; no mastering chain. |
| DAWproject (Studio One, Cubase, Bitwig) | Same as Ableton. | Same as Ableton. |
| Pro Tools | Imports the stems into a running Pro Tools session (current or new), as raw stems or as Auto-Mix processed stems with the gain baked into the audio; renames and colours the tracks; saves. | No pan applied, no plug-ins inserted, no presets, no master chain. Requires Pro Tools open with PTSL enabled and the py-ptsl Python package installed (the app tells you if it is missing). |
Presets
The per-plug-in presets the AI generates are written into your session folder under Presets\ (Gate, EQ, Compression, DeEsser, Saturation, StereoImager, MixBus, Limiter). Exported sessions reference them from there. The AI panel's Engine Presets tab can also generate preset banks and Install to FabFilter copies them into FabFilter's own preset folder.
10. The AI Intelligence panel
Open it with the AI icon in the title bar (Esc closes it). Four tabs.
Analysis
- Run Critique plus a Coach chat box for free-text questions about the mix.
- Run Diagnostics, Analyze Arrangement, Analyze Bounce.
- Spectral heatmap with Fix Masking, which generates bell-cut EQ presets for masking conflicts.
- Analyze Vocal with a stem picker and 11 transcription languages (English, Spanish, French, German, Italian, Portuguese, Japanese, Korean, Chinese, Arabic, Hebrew).
- Detect Key & Tempo per track or for the whole session, and ⚡ Generate All.
- Reference Track Matching: load a reference and your mix, Analyze & Match, choose an Amount (0 to 100%) and Precision, tick Apply match on render, and Preview Match renders a 15-second excerpt. A real-time match meter with target select, integration reset and a freeze overlay sits in the same tab.
Engine Presets
Generate track presets or a 54-preset bank, per-type presets (Saturation, Multiband, De-Esser, Gate, Limiter, Compression), Show Chain Order, and Install to FabFilter.
AI Tools — Music Generation (Labs)
Experimental (Labs). This pane appears after you switch on Settings → Labs → Make a song. Engine select (ACE-Step, with a second engine slot). ACE-Step controls: Auto Setup, Set Path, Start / Stop Server, settings (install path, API URL, default http://localhost:8001, split stem count, auto-start). Generation: vibe prompt with Auto-caption, Instrumental / Vocal / Accompaniment modes, tags, a lyrics generator with history, quality Turbo / SFT / XL 4B (weights download on first use), 1 to 4 variants, BPM, key, duration 5 to 300 s, seed, Auto-fill from Session, Generate, Import to Pro Tools, Cover (source audio plus style prompt and strength, auto-split into stems), and Analyze Generated Stems, which feeds the result straight into a mix session. A phrase library lives under Browse Phrases.
Profiles
Import a portable mix profile, or save the current one.
11. Settings
| Group | What is there |
|---|---|
| Language | The interface language: Automatic (your system language) or any of 25 languages. Applies after a restart (Restart now). The plug-ins follow the same setting the next time a plug-in window opens. Legal documents stay in English; knob, mode and parameter names in the plug-ins stay in English so DAW automation and sessions work in every language. |
| Audio | Output device. |
| GPU and models | Detected GPU; the LLM Model ladder (auto-picked by video memory, section 13); Use your own model: any installed Ollama model, a local .gguf file, or an OpenAI-compatible endpoint (URL, model name, optional API key) with a Test connection button; per-type Model Paths. |
| Dependency paths | Python, ACE-Step and CUDA locations, each with Browse and Clear. |
| Session Storage | Where sessions live, with a Clean Up Old Location button after a move. |
| Plugin Set | Auto, FabFilter (choose Pro-Q 3 or 4 and Pro-C 2 or 3) or StemDesk Built-in. This decides what Auto-Mix renders with and what the preset generators and DAW exports target. |
| Plug-in scanning | Rescan Plugins, Run full plugin load probe, and Custom VST3 Scan Directories. |
| Mix defaults | Default genre and default intensity. |
| Channel Strip Server | Status with Start / Stop / Restart. This is the link on port 9800 that every plug-in uses. |
| Appearance and logging | Light Mode; Debug Logging. |
| License | Tier, email, Manage License. |
| Updates | Check for Updates opens a dialog with the new version's notes and a Download button. Updates are never installed automatically. |
| Diagnostics | Host-shim log path (Copy path / Reveal in folder), View DAW Compatibility, First session guide, Save Diagnostic Report, Reset to Defaults. |
12. Keyboard shortcuts
| Keys | Action |
|---|---|
| Ctrl+S | Save session |
| Ctrl+Z / Ctrl+Y | Undo / redo gain, pan, intensity, role and group changes |
| Ctrl+Shift+A | Select all tracks |
| S / M | Solo / mute the track under the mouse |
| Space | Play / pause |
| Tab | A/B toggle |
| Ctrl+F | Focus the track filter |
| Esc | Close the filter, About dialog, AI panel or Settings |
| Ctrl+Shift+L | License dialog (Ctrl+Enter inside it activates) |
Use Cmd in place of Ctrl on macOS.
13. Platform notes and system requirements
- Disk. About 3 GB for the app on Windows, plus the AI model (2.8 to 30.4 GB depending on your GPU), plus separation and generation weights on first use.
- AI model ladder. StemDesk chooses among 10 local models by available video memory, from a 2.8 GB model for a 4 GB card to a 27 GB near-lossless model for 46 GB and up. Smaller models give shorter, less detailed critiques. On a multi-GPU machine the memory is pooled. You can override the choice in Settings.
- Music generation needs an NVIDIA GPU with CUDA locally, or a remote ACE-Step server.
- Port 9800 on the loopback interface must be free and not blocked by security software; every plug-in feature that says "needs the app" uses it.
- Windows ships the plug-in manager helper that installs and updates plug-ins; it is Windows-only.
- Your files on Windows. Each Windows user has their own StemDesk folders:
Documents\StemDesk(sessions, models, logs),%APPDATA%\StemDesk(settings and license) and%LOCALAPPDATA%\StemDesk(the AI runtime). Uninstalling removes them only for the user who runs the uninstaller; another user on the same PC can delete their own folders by hand. - macOS is Apple Silicon only. VST3, AU and AAX. Python dependencies install on first run.
- Linux is an x86-64 AppImage. VST3 only. Pro Tools export is unavailable because Pro Tools does not run on Linux. Verify the download with the published GPG key, fingerprint
0213 8837 D626 85A4 0CA4 1265 D36C 257A C65B A130.
14. Troubleshooting
| Symptom | What to do |
|---|---|
| Strip says APP OFFLINE or AI offline while the app is open | Settings → Channel Strip Server → Restart. Check that nothing else is using port 9800 and that security software allows loopback connections. |
| Your DAW does not show the paid plug-ins | Activate the license first, then Settings → Rescan Plugins, then restart the DAW. On Windows the app installs them through its helper; give it a moment after activation. |
| ANALYZE shows NO AUDIO | Play the track through the strip while it captures. Some hosts do not feed plug-ins while stopped. |
| ANALYZE shows FAILED with "capture is silent" | Audio reached the strip but at silence level. Check routing and input trim, then try again. |
| Wrong instrument detected | Capture a more representative section and press ANALYZE again. In Workflow B, rename the stem file to its instrument. |
| Auto-Mix, exports or Batch are locked | They are Studio features. The lock badge opens the upgrade dialog. |
| Mix export items are greyed out | Run Auto-Mix first. Master items need a mastering target on that run. |
| AI panes do nothing | The model download has not finished or was declined. Settings → LLM Model, or the amber banner on the Generate pane. |
| Generate pane is inert | ACE-Step is not installed (no NVIDIA GPU) or its server is not running. Start it under AI Tools, or point the API URL at a machine that has it. |
| Pro Tools export fails | Pro Tools must be running with PTSL enabled, and py-ptsl must be installed for the bundled Python. |
| Exported Ableton or DAWproject session sounds raw | Expected: those formats carry fader and pan only, plug-ins at default state. Use the REAPER export in FabFilter mode for an embedded mix, or keep StemDesk running so the strips are configured live. |
| Windows SmartScreen warning | Publisher is Fulks, Inc. More info → Run anyway. |
15. Getting help
Settings → Save Diagnostic Report writes a report with no audio and no file contents; your user name, email, license key, session names and audio file names are redacted where the app can recognise them, before it is saved. Email it, with a short description of what you did, to support@stemdesk.ai. For plug-in render questions, support may also ask for the host-shim log, whose path is shown next to that button.
StemDesk v1.0.53 · Fulks, Inc. · STEMDESK™ is a trademark of Fulks, Inc. · stemdesk.ai · Release notes