FXStudio

FX Studio

Create sound effects locally. Let your coding agent create them too.

FX Studio is a desktop app for sound effects powered by Stable Audio 3 Medium. Describe a camera shutter, UI click, transition or game action, generate several takes, audition them, trim the waveform and drag the result into your editor.

A local agent can create a project, generate a named sound pack and export the finished assets through MCP, the local API or an installable skill. Its sounds appear in the same app, ready for you to review.

Built by WIK+DEV · X / @wik_plus · Buy me a coffee

Latest source updates: regenerate a sound in place with a fresh seed, choose OGG/WAV/MP3 directly in Generate, and delete projects with their internal sound files. See the changelog for details and validation.

Screenshots show real Medium output in a separate demo project.

Download and install

Get the compiled app from GitHub Releases. GitHub's Code → Download ZIP contains source code, not an installed app.

The latest changes on main are listed as Unreleased in the changelog. Existing v0.1.0 release downloads do not include them; build from source or use a newer release once published.

Platform	Download	Generation engine	Status
macOS 26.2+, Apple Silicon	FXStudio-macOS-arm64.zip	Official MLX	Tested on an M1 Pro with 16 GB unified memory
Windows 10/11, x64	FXStudio-Windows-x64-Setup.exe or portable ZIP	Official LiteRT, CPU	Native build, installer and real Medium generation tested

Intel Macs, Windows ARM and Linux are not supported by these packages. Windows uses quantized Medium weights; its speed and output are not identical to MLX. A small test clip is not a promise of performance on every computer.

macOS

This package requires macOS 26.2 or newer because of its bundled MLX native libraries. The minimum is read from the actual bundled binaries, rather than inherited from Electron.

1. Download and unzip FXStudio-macOS-arm64.zip.
2. Move FX Studio.app into Applications.
3. Open it. If macOS blocks the unsigned developer, try System Settings → Privacy & Security → Open Anyway, then confirm the app you downloaded. See Apple's instructions. The app is not notarized.
4. Let the first-launch model download finish. You can pause or retry it inside the app.

No Terminal, Homebrew, Node.js or separate Python installation is needed to use the compiled app.

Windows

The bundled LiteRT runtime requires an x64 CPU with AVX2, FMA and F16C, matching its official Windows build flags. Windows ARM is not supported by this package.

Install the latest Microsoft Visual C++ v14 Redistributable, x64 if it is not already present. The official LiteRT wheel needs this Microsoft runtime, including VCRUNTIME140_THREADS.dll. The FX Studio installer links to Microsoft's download page when the runtime is missing.

Download the Setup EXE from a release that includes a tested Windows asset. Run it to install for your user; administrator rights are not required. The portable alternative is FXStudio-Windows-x64.zip: extract the whole folder, then open FX Studio.exe. Keep resources and the other files beside it.

Windows packages are unsigned. Check the release's checksum and download source before deciding whether to allow a Windows reputation warning. Do not disable your computer's protection globally.

Native Windows CI passed 62 behavior tests, standalone MCP/CLI/project and crop/MP3 checks, silent installation/uninstallation, and a real one-second Medium generation. Desktop playback and dragging into a particular editing application still need hands-on testing on Windows.

The model downloads automatically

Only Stable Audio 3 Medium is used. Weights are downloaded separately from Stability AI's pinned Hugging Face release when the app first opens. Setup shows real download progress, missing files, disk space, pause/resume and errors. Existing complete files and the local cache are reused. After setup, generation works offline.

Model download	Disk space for weights and text encoder
Mac / MLX	6.41 GiB (6.88 GB)
Windows / LiteRT	about 2.99 GiB, quantized w8a8

These are download/storage sizes, not measured RAM requirements. Leave additional space for the app, temporary downloads and generated audio. If upstream access requires authorization, the app reports the actual access error; it cannot accept agreements for you.

The application source is MIT licensed. The downloaded model has separate Stability AI Community License and Gemma terms. Review model and third-party terms before use; “open-source app” does not mean the model weights are MIT licensed.

What you can do

• Generate 0.5–30 second stereo sound effects at 44.1 kHz, with 1–4 variants.
• Create projects for a game, video or asset pack. Review takes with Previous/Next, favorites and search.
• Play, seek and zoom the real waveform. Drag either edge to trim; crops get automatic 5 ms fades. The original audio remains available.
• Choose Generate → Format: smaller OGG Vorbis (quality 5, initially selected), lossless WAV or 320 kbps MP3. The choice persists across restarts and becomes the default for agent generation; agents can override it with format: "ogg", "wav" or "mp3". Each queued job keeps its selected format. Regenerate preserves the sound’s format.
• Drag the saved crop into an editor, or export WAV, OGG Vorbis (quality 5) and 320 kbps MP3. OGG creates smaller lossy delivery files from the saved crop; new generations are stored in the chosen format. Compressed storage removes the engine’s temporary WAV after a successful save. Dragging a crop keeps the stored format. Existing sounds retain their format. Agent exports default to format: "ogg"; WAV/MP3 conversion remains available. Converting OGG to WAV does not restore the discarded detail. Stable drag-export files survive app restarts and later trim changes.
• Make variations from an existing sound or regenerate a selected section.
• Click Regenerate on a generated sound for a fresh seed with the same prompt and settings. The new take replaces it in place, retaining projects, favorites and its saved crop. You can cancel; failed or cancelled attempts keep the previous take.
• Queue jobs, see actual progress/errors and cancel generation. One worker runs at a time and releases its model memory when it exits.
• Use the same workflow from local agents, with named packs, project membership and export manifests.

Sounds, projects, seeds, settings, trims and favorites persist locally. In Projects, Delete project previews and permanently removes its exclusive generated sounds, previous takes, crops, internal drag files and job logs after confirmation. Sounds shared with other projects, the sound library and external exports are preserved. Finish or cancel affected jobs before deletion. Local agents can use preview_project_deletion and delete_project for explicitly requested deletion. There are no subscriptions, purchase gates, public generation endpoints or cloud agent connections.

Get cleaner results

Start at 8 steps, CFG 1, then make 2–4 variants and listen. More steps and higher CFG do not automatically improve a take. At CFG 1, a negative prompt has no effect.

Use one concrete event per asset: source → action/event count → texture and decay → recording perspective and room. For close foley, specify a dry recording. Match the requested duration to the event, then trim the actual result. These original examples follow the official Stable Audio 3 prompt guide:

Camera shutter — 1 second

A DSLR camera shutter firing once. A crisp mechanical snap followed by a tiny spring click, with a compact midrange body and a very fast decay. Recorded close to the mechanism in an acoustically dead room.

UI click — 1 second

A single soft synthesized UI tap. A rounded midrange tick with a delicate bright attack and a very short, smooth decay. Dry and isolated.

Transition — 2 seconds

A quick airy whoosh sweeping past the microphone once. A smooth rising rush of air with a soft low body, a bright peak and a short clean tail. Close, dry studio sound design.

Wooden door — 2 seconds

A small wooden cabin door pushed shut once. A compact wooden knock followed by a short latch click, with a warm low-mid body and a quick natural decay. Recorded close in a quiet, dry room.

Page turn — 2 seconds

One page of a thick paper notebook turned slowly. A light papery scrape, a brief rustle and a soft settling tap. Recorded very close in a quiet room with little reverberation.

More examples are included in the app and English agent skill. A fixed seed helps compare changes; a blank seed creates fresh takes. Audition the results: prompts cannot guarantee an exact event placement or perfect sound quality.

Use with a local agent

1. Install/open the app. In Setup → Local agents, enable access and choose an export folder.
2. Click Install skill… → Codex, or choose another local agent's skills folder. Start a new agent conversation. Save agent kit… exports the skill, helper and configuration recipes for your current installation.
3. Ask the agent, for example:

Use FX Studio to create a project called Forest game. Generate two variations each of a soft UI select, a wooden door closing and a player landing on grass. Keep the sounds dry and short. Export the selected assets as WAVs and tell me which takes you were able to audition.

The skill includes concrete prompts, settings, project creation, polling, cancellation, trimming and exports. It supports a command-line fallback for agents that can run local commands. No extra Python install is required: the app includes its own interpreter. Agents must be able to access this computer; a skill alone does not give a cloud agent access to your Mac or PC.

MCP

In Setup, Copy MCP config gives standard mcpServers JSON; Copy Codex config gives TOML for ~/.codex/config.toml. Paste the generated configuration into the local client's MCP settings, then reload it. Paths are specific to where you installed the app. The stdio adapter reads its local private credential automatically; never paste an API key into a prompt or public configuration.

There are 22 tools: status and prompt guide; sound/pack generation, polling and cancellation; history/library search; sound inspection, trimming, favorites and exports; project creation, updates, sound membership and confirmed project deletion. The guide is also exposed as fxstudio://guide.

Agent-created jobs appear in the desktop queue. Pack creation returns immediately; poll the returned IDs. Finished takes are saved in the selected project, and exported packs include a manifest with prompts, settings, seeds, trims and provenance.

API / CLI

The authenticated local API is at http://127.0.0.1:3921/agent/v1. Access starts off, can be disabled or reset in Setup, and exports stay within the folder you selected. The service binds only to loopback. Generation does not have a short fixed timeout.

For complete routes, examples, CLI commands and client setup, read AGENT-ACCESS.md. The installable skill source is ui/skills/fx-studio.

Your data

Platform	App data
macOS	~/Library/Application Support/FX Studio/
Windows	%LOCALAPPDATA%\FX Studio\

Generated audio is in audio/; projects/history/favorites/trims are in studio.sqlite; crop caches are in edits/; stable editor references are in drag-exports/; logs are in jobs/ and service.log. Model links/files are in models/. Keep drag-export files while your editor still references them. Removing/reinstalling the app does not remove this data folder.

No personal sound library is included. An optional catalog can be connected using FX_LIBRARY_ROOT or an absolute path in library-root.txt in the app data folder. Without configuration the app looks in ~/Music/FX Studio Library. It expects katalog.json with existing assets[].path entries and provenance; it reads those files in place and retains documented/unknown rights. Generated sounds work independently of this optional library.

Build from source

End users should use Releases. Developers need Git, Node.js/npm and uv.

git clone https://github.com/WikPlus/FXStudio.git
cd FXStudio

Apple Silicon Mac: install Xcode Command Line Tools. ./Einrichten.command prepares the pinned official MLX runtime, Python dependencies, Electron and Medium weights; npm start launches development mode. To build a standalone release, install pkg-config and shared LAME 3.100, then:

npm run prepare:release
npm test
npm run check
npm run build
npm run package:mac

The collector requires the bundled upstream license materials and exact pinned native versions. Read release engineering before rebuilding native libraries.

Windows x64: the included GitHub Actions workflow provisions the build tools, compiles native audio components, installs the official LiteRT runtime, runs behavior/runtime/installer tests and uploads tested packages. For a local build, install MSYS2 at C:\msys64 with the packages listed in the workflow, and Inno Setup 6, then run ./Build-Windows.ps1 in PowerShell. A custom native build must supply matching license/source evidence.

Normal CI does not download model weights. A manually dispatched run can explicitly enable a temporary real Medium generation test. Test fixtures create disposable synthetic audio to check behavior; they are not presented as generated examples. CI reports distinguish runtime/export checks from actual model inference.

License and support

MIT for FX Studio's original code, documentation and agent skill. See NOTICE and THIRD_PARTY_NOTICES.md for bundled software, model terms and corresponding native-library sources. Setup can open all bundled licenses offline.

Report reproducible problems in Issues, including platform, app version and the error. Remove prompts, personal paths and credentials from shared logs. Contributions are welcome; test the behavior you change and keep model weights, user audio and secrets out of Git.

If it helps your workflow, you can buy me a coffee.