Local AI media creation studio for macOS, powered by Apple Silicon. Generate images, videos, audio, and music -- all locally. Multi-backend image generation (Automatic1111, ComfyUI, SwarmUI, MLX), 5-backend LLM chat, a full audio suite with voice cloning, and a WidgetKit extension.
Written by Jordan Koch (@kochj23).
graph TB
subgraph App["AI Studio (SwiftUI macOS)"]
Images[Images Tab] & Videos[Videos Tab] & Audio[Audio Tab] & Chat[Chat Tab] & Gallery[Gallery Tab]
end
subgraph ImageBackends["Image Generation"]
A1111[Automatic1111\nREST API]
ComfyUI[ComfyUI\nREST + WebSocket + ControlNet]
SwarmUI[SwarmUI\nREST Sessions]
MLXImg[MLX Image\nPython Daemon]
end
subgraph LLMBackends["LLM Chat"]
Ollama[Ollama\nStreaming]
TinyLLM[TinyLLM\nSSE]
TinyChat[TinyChat\nREST]
OpenWebUI[OpenWebUI\nSSE]
MLXLLM[MLX Native\nSubprocess]
end
subgraph Daemon["Python Daemon (stdin/stdout JSON)"]
daemon[aistudio_daemon.py]
daemon --> tts[mlx_tts.py\n6 TTS Engines]
daemon --> voice[mlx_voice_clone.py\nf5-tts-mlx]
daemon --> whisper[mlx_whisper_stt.py]
daemon --> music[mlx_music_gen.py]
daemon --> img[mlx_image_gen.py]
end
subgraph Services["Core Services"]
BackendMgr[BackendManager\nHealth checks + selection]
LLMBackendMgr[LLMBackendManager\nAuto-detect + fallback]
Queue[GenerationQueue\nFIFO, 50 items]
History[PromptHistory\nPersistent, search, tags]
Retry[RetryHandler\nExponential backoff]
end
Images --> BackendMgr --> ImageBackends
Chat --> LLMBackendMgr --> LLMBackends
Audio --> daemon
Videos --> ComfyUI
Queue --> BackendMgr
ImageBackendProtocol abstracts all image backends behind an actor-based interface. The Python daemon communicates via stdin/stdout JSON-line protocol with UUID request IDs for concurrent requests. Modules are lazy-loaded on first use.
| Backend | Protocol | Highlights |
|---|---|---|
| Automatic1111 | REST | txt2img and img2img |
| ComfyUI | REST + WebSocket | Workflow-based, ControlNet support |
| SwarmUI | REST | Session-based generation |
| MLX Native | Python daemon | On-device via mflux/diffusionkit |
Model picker with auto-detection per backend. SafeTensors-only enforcement (.ckpt, .bin, .pt blocked). Parameter controls: steps, CFG, sampler, dimensions, seed, batch size. Auto-save with date-organized output and metadata JSON export.
Batch up to 50 prompts with FIFO processing, pause/resume, reorder, and cancel. Persistent prompt library with search, tags, favorites, deduplication, and one-click reload.
Side-by-side mode with synchronized zoom. Slider overlay mode with draggable divider. Fit, 1:1, and zoom controls.
AnimateDiff via ComfyUI with frame-to-MP4 combining (AVAssetWriter). Configurable frame count, FPS, and resolution.
- TTS -- 6 MLX engines via mlx-audio: Kokoro (11 voices), Dia, Chatterbox, Spark, Breeze, OuteTTS
- Voice Cloning -- f5-tts-mlx with auto-transcription via mlx-whisper and automatic sample rate conversion (24kHz)
- Speech-to-Text -- mlx-whisper (tiny through large-v3)
- Music Generation -- MusicGen via transformers
6 backends (Ollama, TinyLLM, TinyChat, OpenWebUI, MLX, OpenRouter frontier models) plus an optional Nova Gateway, with streaming for the OpenAI-compatible backends. Auto-detection with health-checked priority-based failover. Conversation history, configurable system prompt, temperature, and max tokens. OpenRouter keys are stored in the macOS Keychain, never in plain settings.
Beyond single-backend failover, AIStudio can spread chat requests across every model available to it — the single-user version of how Nova's gateway balances load. A ModelRegistry discovers models across backends and a pure, network-free LoadBalancer (round-robin or least-busy) picks the next one, health-gated so unhealthy models are skipped and it falls through cleanly. Three toggles in Settings → LLM control the pool:
- Use all local models — every discovered Ollama + MLX model on this Mac joins the balancer pool.
- Enable all frontier models — OpenRouter's full model list joins the pool (requires a stored OpenRouter key).
- Route through Nova Gateway — registers a "Nova Gateway" backend (OpenAI-compatible, default
127.0.0.1:18792) that the app routes to, inheriting Nova's own internal routing.
With every toggle off, behavior is unchanged (classic single-backend + failover).
Browse all generated media. Filter by type, search by prompt, sort by date. Metadata panel with full generation parameters. Reveal in Finder and delete.
Small, medium, and large widgets with backend status and recent generation info via App Group.
Retry with exponential backoff and jitter (3 presets: httpBackend, pythonDaemon, healthCheck). Python daemon crash recovery with auto-restart up to 5 times.
Port 37425, loopback only. GET /api/status and GET /api/ping.
- Download the latest
.dmgfrom Releases. - Open it and drag AIStudio into your Applications folder.
- Launch it from Applications.
See "AIStudio can't be opened because the developer cannot be verified"? That means you have a build that isn't yet Developer-ID-signed and notarized. To open it anyway:
- macOS 14 and earlier: Control-click (right-click) the app → Open → Open.
- macOS 15 (Sequoia) / 26 and later: double-click it, dismiss the dialog, then open System Settings → Privacy & Security, scroll down, and click Open Anyway.
- Or from Terminal:
xattr -dr com.apple.quarantine "/Applications/AIStudio.app"Notarized releases open with no prompt at all — maintainers, see RELEASE.md.
Requires macOS 14.0 Sonoma and Apple Silicon. Distributed via DMG only (no Mac App Store). Sandbox disabled for file system access and Python subprocess management.
Requires Xcode 15 or later (macOS 14 SDK). AIStudio's MLX-native backend runs through an embedded
Python runtime (mflux / mlx-audio / mlx-whisper), not the mlx-swift Swift package, so the
Xcode build compiles no Metal shaders and no Metal Toolchain is required to build.
git clone git@github.com:kochj23/AIStudio.git
cd AIStudio
open AIStudio.xcodeproj # Build & run: Cmd+RThe MLX-native features are optional and configured at runtime — see MLX Native Setup below.
python3 -m venv venv && source venv/bin/activate
pip install 'mlx-audio[kokoro]' f5-tts-mlx mlx-whisper numpy Pillow
brew install espeak-ng # for voice cloning phonemizer
pip install mflux # for MLX image generationSet the Python path in Settings to your venv's python3 binary.
- SafeTensors-only -- PyTorch pickle files blocked (supply chain attack vector)
- Prompt injection prevention -- User prompts written to temp files, not embedded as string literals
- stdout isolation -- ML operations redirect stdout to protect JSON protocol
- SecureLogger -- Redacts API keys, tokens, and PII from all log output
- Path traversal prevention -- Validated against
../, symlink resolution, 4096-byte cap - No cloud services, no telemetry, no analytics -- All communication is localhost
205 Swift + 65 Python = 270 total tests.
| Suite | Tests | Coverage |
|---|---|---|
| ComprehensiveTests | 78 | Cross-module integration, end-to-end workflows |
| ModelTests | 45 | All data models, enums, Codable, error types |
| SecurityUtilsTests | 25 | Input validation, path traversal, URL/HTML sanitization |
| SecureLoggerTests | 18 | PII redaction, API keys, JWTs, phone numbers, credit cards |
| FileOrganizerTests | 14 | Date directories, filename sanitization, metadata export |
| ImageUtilsTests | 13 | PNG/JPEG validation, base64, file size limits |
| RetryHandlerTests | 12 | Backoff, jitter, presets, cancellation |
| Suite | Tests | Coverage |
|---|---|---|
| test_tts | 21 | Engine config, voice listing, WAV encoding |
| test_daemon | 20 | JSON protocol, request dispatch, input safety |
| test_music_gen | 7 | WAV encoding, model caching |
| test_voice_clone | 7 | WAV encoding (24kHz), resampling |
| test_image_gen | 5 | Pipeline init, model listing |
| test_whisper_stt | 5 | Model caching, standard sizes |
| Version | Date | Highlights |
|---|---|---|
| 2.4.0 | Aug 2026 | Multi-model load balancing (ModelRegistry + LoadBalancer), 3 pool toggles, Nova Gateway backend; OpenRouter frontier models + Keychain key + health-checked failover |
| 2.3.2 | Mar 2026 | Prompt injection fix, SafeTensors enforcement, URLComponents safety |
| 2.3.1 | Feb 2026 | Daemon pipe buffering fix, voice cloning auto-transcription, TTS rewrite for mlx-audio 0.3.x |
| 2.3.0 | Feb 2026 | Queue, prompt history, image comparison, ControlNet, LLM streaming, retry/backoff |
| 2.2.0 | Feb 2026 | WidgetKit extension |
| 2.1.0 | Feb 2026 | LLM chat with 5 backends |
| 2.0.0 | Feb 2026 | Complete rewrite: audio suite, video, gallery, multi-backend |
| 1.0.0 | Feb 2026 | Initial release with Automatic1111 |
MIT License -- see LICENSE.
Copyright (c) 2026 Jordan Koch. All rights reserved.
Written by Jordan Koch (@kochj23).
Disclaimer: This is a personal project created on my own time. It is not affiliated with, endorsed by, or representative of my employer.