The ElevenLabs sound generator, without leaving your DAW.
VST3 · AU · standalone — free, and you bring your own
ElevenLabs key.
The builds are unsigned: macOS needs
one command before it will load them.
ElevenLabs makes the sound effect. PromptSlice is the VST3 / AU plugin you work with it in: ask for a handful of takes from inside the project, hear them against the track that is already playing, mark up the one that worked, cut the piece you want out of it and drag it onto a channel.
What it removes is the round trip. A browser tab, a download folder, a file manager, and all of it again for the next take — for a sound you will keep two seconds of.
VST3 and AU, so it loads anywhere that takes them — Ableton, Logic, Reaper, Bitwig, Cubase, Studio One. It was written and tested in FL Studio 2026, and that is the one host where the last step is proven: dragging a cut into the project is a file drop the host has to accept, and hosts differ. If yours does something else, that is worth an issue.
| Several takes at once | A prompt gives one sound, so asking for five is five requests — and the first plays while the rest are still arriving. |
| The whole editing bit | Playhead, markers, a selection, looping, cutting. Cuts land on the nearest zero crossing within 10 ms and the last millisecond fades, so a piece does not click at either end. |
| Lossless | Takes are requested as PCM rather than the endpoint's default of 128 kbps mp3, which smears exactly the transients a one-shot is made of. |
| Honest about money | Every request reports what it cost, and the plugin shows that figure rather than guessing against a price list. |
| Nothing is lost | What each batch was asked for is written beside the sounds it made, and Library opens it again — prompt, settings and all. |
There isn't one. VST3 gives a plugin no way to create a channel, place a clip, or write anything into the host's project — that is not an omission here, it is the format. What a plugin can do is write a file and let the host accept a drop of it, which is how Serato Sample, XO and Portal all do it too. So every piece you cut lands in the row at the bottom, and you drag it where you want it.
Built releases are on the releases page — a zip per system, with where to put things and what macOS will say about it.
Important
Those builds are not signed or notarised, which needs a paid Apple developer account. macOS will refuse to load a downloaded plugin until you clear the quarantine mark by hand; the release notes give the command. If you would rather not take that on trust, build it yourself — it is three lines.
You need CMake 3.22+ and a C++20 compiler. JUCE is fetched automatically
at the version this was tested against — if you already have a checkout, add
-DJUCE_PATH=/path/to/JUCE and save the download.
Both platforms are built on every push by CI, so what follows is what a machine does, not what somebody remembers doing.
Xcode command line tools · macOS 11 or newer
git clone https://github.com/vasylNaumenko/promptslice.git
cd promptslice
cmake -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo
cmake --build buildBuilds VST3, AU and a standalone app, universal (arm64 and x86_64) so the plugin
is visible however the host was started, Rosetta included. It copies itself into
~/Library/Audio/Plug-Ins/; rescan plugins in your DAW afterwards.
Visual Studio 2022 or newer with Desktop development with C++ · Git
git clone https://github.com/vasylNaumenko/promptslice.git
cd promptslice
cmake -B build -A x64
cmake --build build --config RelWithDebInfoCMake picks whichever Visual Studio it finds, so there is no version to keep up
to date here. Naming one with -G is possible and rarely what you want: it
fails outright on a machine that has a different one.
Two differences from the macOS build, both of them JUCE's doing
No AU. Audio Units are an Apple format; JUCE drops that target on any other platform. You get VST3 and the standalone.
It does not install itself. VST3 plugins go in
C:\Program Files\Common Files\VST3, which an ordinary build cannot write to —
and the copy is a post-build step, so it would not be skipped, it would fail the
build. Copy build\PromptSlice_artefacts\RelWithDebInfo\VST3\PromptSlice.vst3
there yourself (it is a folder, not a file — copy the whole thing), or from an
elevated prompt let the build do it:
cmake -B build -A x64 -DPROMPTSLICE_INSTALL_AFTER_BUILD=ONOn macOS that switch is on by default, because there the destination is inside your own home folder.
PromptSlice is an instrument, not an effect: an instrument track in most hosts, the Channel Rack in FL Studio. Type what the sound should be, press Generate, and the takes arrive in the row along the bottom. The first one plays while the rest are still coming.
Everything you can drag into the project is there. A chip carries its file's own name rather than a position, so deleting one cannot renumber the rest, and the newest is always at the end. The colour says where it came from:
| 🟦 Blue | a take, generated from the prompt |
| 🟪 Purple | a cut, taken out of one of them |
Press a chip to hear it, drag it into the project, or press its ✕ to send it to the trash.
New lets go of the batch you are on: the prompt and the row clear, the settings you were working with stay. It touches no files.
Library opens a batch you made before — its sounds and the settings that produced them, read back from the note described below.
Mouse and keyboard
| Click the waveform | move the playhead, drop the selection |
| Drag the waveform | select a stretch |
| Drag the bar on top of a selection | cut it and drag it straight out |
| Double click | add a marker |
| Right click a marker | remove it |
| Drag a marker | move it, bounded by its neighbours |
| Wheel / ⌘+wheel | scroll / zoom about the pointer |
Space, M, Esc |
play, marker, drop the selection |
Cut slices the stretches between markers — two markers mean one piece. Markers win over a selection, and the button says which it will do.
Everything lands under the folder in the settings, one subfolder per batch, named after the prompt and stamped. Change it with Save to…, or edit the settings file directly:
macOS ~/Library/Application Support/PromptSlice/config.json
Windows %APPDATA%\PromptSlice\config.json
It lives in that file and nowhere else — never in a project, never in this
repository. Set it with Key…, or put it in api_key there. Keys are made at
elevenlabs.io/app/developers/api-keys,
where each one is given the access it is allowed to use.
PromptSlice calls three endpoints and no others, so it needs two things and nothing else:
| Access | What it is for | |
|---|---|---|
| Sound generation | required | POST /v1/sound-generation. Without it nothing works at all. |
| User Access → Read | optional | Reading /v1/user/subscription for the real balance. Without it the plugin falls back to what the key has spent, and says so. |
Everything else can stay off — History, Models, Voices, Text to
Speech, Dubbing, the lot. History is the obvious wrong guess and worth
naming: it is your library of past generations on the site, and the plugin never
opens it. Measured rather than assumed — a key with models_read and
voices_read refused runs this plugin without noticing.
Each batch folder gets a promptslice.cfg — plain JSON, meant to be readable,
saying what that batch was asked for:
{
"prompt": "deep metal impact with a long reverb tail",
"takes": 3,
"length_seconds": 2.5,
"prompt_influence": 0.75,
"model": "eleven_text_to_sound_v3",
"sample_rate": 48000,
"credits": 75
}That file is why a folder of wavs stops being anonymous a week later, and it travels with the folder if you move or copy it. Edit it by hand if you like — what it says is checked on the way back in, and anything it leaves out comes back as a default.
| Model | Auto names no model and lets ElevenLabs pick. Sound v2 and Sound v3 ask for one by name. |
| Quality | The rate the wav is written at, 22.05 to 48 kHz. |
Neither changes the price: the service charges by the second of audio, so the rate trades disk space and nothing else.
Every request comes back saying what it cost, and that figure — the service's own, not a guess — is what the status line reports:
Generated 3 takes for 60 credits.
Beside it the plugin says what it can about the account, and what that is depends on what the key is allowed to read:
| The key can read | It shows |
|---|---|
the account (user_read) |
12 340 credits left — the real balance |
| only its own usage | 1124 credits spent in 30 days |
| neither | balance unavailable, with the reason in the tooltip |
The middle row is what a key scoped to sound generation gets, and it is worded as
what it is: spent is not left. To get the real balance, turn on the
user_read permission where the key is made.
Tip
A Length of auto is charged as a flat rate rather than by the second, and
for a short sound that is markedly more than naming the length. Measured
against the live endpoint: half a second asked for costs 5 credits, two seconds
costs 20 — ten a second — while auto came back as one second of audio for 50.
cmake --build build --target SliceTest && ./build/SliceTest_artefacts/*/SliceTestSliceTest checks what happens to files you keep — the zero-crossing snap, the
fade, the numbering, the refusals, the order the row puts them in, and the
settings prune, which is the only code here that deletes anything. It needs
nothing external and runs against a synthetic tone.
GenTest covers the batch note the same way, then makes one real request and
checks the answer opens as a wav at the rate it asked for, that the cost was
reported, and that the balance query gives either a figure or a reason. The note
half runs always; the network half skips when there is no key, so it is safe to
run in any state:
ELEVENLABS_API_KEY=... ./build/GenTest_artefacts/*/GenTestAGPL-3.0-or-later. See LICENSE.
This is not a preference: the plugin links the JUCE framework, which is dual-licensed under AGPLv3 or a paid commercial licence. A build made without a commercial JUCE licence is a derivative work and can only be distributed under AGPLv3. In short — use it for anything you like, but if you distribute a modified version, its source has to be available under the same terms.
ElevenLabs, with your key, to ask for sounds and to ask what they cost. Nothing else leaves the machine: no telemetry, no analytics, no update check.
Sounds you generate are yours to use under ElevenLabs' own terms, which are between you and them.