Agave Studio

One studio for AI image, video, voice and 3D, with a video editor built in.
Free and open source. Bring your own API keys.

There are no plans, credits or accounts. You add a key from a provider such as fal.ai and pay them directly for what you make. For fal.ai models the studio shows the provider's list price before you press Generate (Meshy jobs show their credit cost). Keys are stored only on your computer, and the app has no analytics.

What it does

• Generate: text to image, image to image, text to video, image to video, voice, and image to 3D, from one queue. gateway/models.toml lists every model.
• Direct: 20 camera moves, 10 looks, a prompt library, and Elements: saved characters or products you call with @tag.
• Edit: trim and join clips, auto captions, remove pauses, reframe to 9:16, clean up voice, and set loudness to -14 LUFS.

Run it

Agave Studio runs on macOS and Linux. You need:

• Rust 1.88 or newer (install it with rustup; distro packages are often older),
• a C++17 compiler (on macOS, the Xcode Command Line Tools: xcode-select --install; on Debian or Ubuntu, build-essential),
• ffmpeg on your PATH.

git clone https://github.com/LayerMakerLab/agave-studio.git
cd agave-studio/gateway
cargo run --release        # the site and engine on http://127.0.0.1:4170

Then open http://127.0.0.1:4170/app/. On first launch the studio walks you through setup: paste a provider key, then download the speech model for captions (a one-time, free download). You can reopen setup from Settings.

The gateway only listens on 127.0.0.1:4170. It has no sign-in, so do not expose it to a network. It reads web/, config/ and data/ from the checkout it was built in, so run it from there.

Providers and keys

Paste keys in the studio's Settings, or put them in config/keys.env (copy config/keys.example.env). Environment variables with the same names also work.

Provider	Key	What it unlocks
fal.ai	FAL_KEY	Kling, Seedance, Wan, Hailuo, Veo, FLUX, Nano Banana, Seedream
Replicate	REPLICATE_API_TOKEN	FLUX schnell
Google Gemini	GEMINI_API_KEY	Veo 3.1, Nano Banana
OpenAI	OPENAI_API_KEY	GPT Image
Black Forest Labs	BFL_API_KEY	FLUX 1.1 pro, FLUX.2 pro
ElevenLabs	ELEVENLABS_API_KEY	Text to speech
Meshy	MESHY_API_KEY	Image to 3D (GLB, USDZ, STL, 3MF)

Every model id and request shape in models.toml was checked against the provider's own docs or API schema; each entry notes what was checked and when. Not every model has had a live run yet, so if one fails, please open an issue.

macOS extras (optional)

These need macOS 14 or later and Xcode's Swift tools.

• Captions with Parakeet (on-device, fast): native/speech/build.sh. Without it, captions use whisper.cpp: put whisper-cli on your PATH and pick Whisper in setup.
• Follow reframe (keeps the subject in frame for 9:16): native/vision/build.sh.
• Mac app: run cargo build --release in gateway/ first, then native/macos/build-app.sh builds native/macos/build/Agave Studio.app, which starts the gateway from this checkout.

Privacy

• Keys stay in config/keys.env on your computer; the studio writes that file with mode 0600 and never sends a saved key back.
• Prompts and files go only to the provider you pick, using your key.
• Captions and transcripts run locally.
• No analytics, no telemetry, no account. The fonts ship with the studio; only the 3D viewer may fetch its decoders from Google's CDN (web/vendor/README.md).
• The data/ folder (outputs, job history) is owner-only on disk. While the gateway runs, any program on this computer can call it on 127.0.0.1; see SECURITY.md.

Layout

Path	What
gateway/	Rust (axum) engine: the API, provider adapters, jobs, rendering with ffmpeg
engine/promptcraft/	C++ camera and style phrases, linked into the gateway
web/	The site (/) and the studio (/app/): plain HTML, CSS and JavaScript, no build step
native/	Optional macOS tools: speech, vision and the Mac app
docs/API.md	The gateway's HTTP API

Develop

cd gateway
cargo test --release       # makes no network calls
cargo clippy --all-targets

See CONTRIBUTING.md and the Code of Conduct. To report a security problem, see SECURITY.md.

License

Apache License 2.0. See LICENSE and NOTICE. Speech models are downloaded from Hugging Face under their own licences: Parakeet TDT 0.6B v3 (CC-BY-4.0) and whisper.cpp's Whisper small (MIT).