A local-first AI virtual camera that turns live speech into matching backgrounds.
Studio. Speech arrives in the rolling transcript, matches a library record at 0.800, and the composited frame is already on its way to the virtual camera.
Aura listens to what you are saying, finds or generates an image that matches it, and composites that image behind you as a real webcam device that Zoom, Teams, OBS and anything else can select. Transcription, matching and rendering all run on your own machine.
Aura has two modes that form a loop:
Record — you capture a short piece of speech. Whisper transcribes it locally, Ollama turns the transcript into a summary, an image prompt and a set of trigger phrases, and ComfyUI generates the background. The result is stored in a local SQLite library.
Play — Aura transcribes your speech continuously and matches it against the trigger phrases in that library, using keyword overlap and Ollama embeddings for semantic matching. Matching images are composited behind you live, with your background segmented out, and the composed frames are published to the virtual camera device.
ARCHITECTURE.md describes the full execution flow, module by module.
The library. Every record keeps its images, a group it belongs to, and the cleaned trigger text that Play Mode matches your speech against.
Aura is Windows-first. Windows is the only platform where every feature works.
| Area | Windows | Linux |
|---|---|---|
| Backend, database, Studio, Play Mode, Slide Output | works | works |
| Segmentation (PaddleSeg) | works | works |
| ComfyUI / Ollama / Whisper integration | works | works |
| Virtual camera device | works | not implemented |
| Prerequisites wizard | works | not ported |
macOS is not supported.
Aura orchestrates several external programs rather than bundling them. It talks to each over local HTTP, so they stay independently installable and updatable.
Required
- Node.js 22.12 or newer — the floor comes from the
electrondevDependency;better-sqlite3itself supports 20.x and up - Ollama — text generation and embeddings. Pull a text model plus
nomic-embed-textfor semantic matching. - Python 3 with the speech service venv — a Flask app wrapping
faster_whisper, listening on127.0.0.1:8765 - FFmpeg — audio handling
Optional
- ComfyUI — image generation. Without it, Record Mode cannot generate new backgrounds, but an imported library still plays.
- Visual Studio with the C++ workload — only needed to rebuild the native virtual-camera components. Prebuilt binaries are committed (see below).
- SerpAPI key — image search import
On Windows, the wizard in installer/prerequisites/ installs and configures
most of the above for you. Run Install Aura Prerequisites.bat and work down
the list.
git clone https://github.com/JeremiahBarkman/Aura.git
cd Aura
npm installThen create your configuration:
cp .env.example .envEdit .env and set the paths for your machine — mainly COMFYUI_* and
SPEECH_*. Every value has a comment explaining it, and the AI service URLs
already default to the standard local ports.
On Windows, scripts/setup-dev.ps1 creates the runtime data folders and checks
that Node is on your PATH:
powershell -ExecutionPolicy Bypass -File scripts/setup-dev.ps1npm start # backend only -> http://localhost:5150
npm run electron # the full appnpm start runs just the Express + SQLite backend. There is no Electron shell,
no virtual camera and no supervised speech or ComfyUI process, which makes it
the fast way to iterate on web, API and database code.
npm run electron runs the real thing: the Electron shell spawns and supervises
the backend, the speech service, ComfyUI, and — on Windows, on demand — the two
native virtual-camera processes.
npm run distThis runs the documentation-sync check, stages the native binaries and the Node
runtime into native/runtime/, then builds an NSIS installer into dist/.
INSTALLER.md covers packaging and deployment in full.
| Port | Service |
|---|---|
5150 |
Aura backend (Express) — the main window loads this |
8188 |
ComfyUI |
8765 |
Speech service (faster-whisper) |
5151 |
Frame bridge, TCP, loopback only |
All of these are bound to localhost.
In a checkout, configuration comes from .env in the repository root and
runtime data is written to data/. Both are gitignored — data/ holds your
SQLite database, your image library and recorded audio, and none of it belongs
in version control.
In a packaged install, Aura relocates both to a per-user directory:
%APPDATA%\aura\ on Windows, ~/.config/Aura/ on Linux. Edit the .env
there, not the one inside the app — the bundled copy is overwritten on every
update, while yours is left alone.
Application settings that the UI owns live in data/config/aura-config.json,
behind a strict key allowlist in src/config/auraConfig.js.
Background segmentation runs in the renderer on a bundled PP-HumanSeg ONNX
model (public/models/paddleseg/, Apache-2.0) through onnxruntime-web, using
WebGPU with a WASM fallback. It needs no credentials, so a fresh clone has a
working camera. The Erode and Background Blur sliders in Studio tune the mask
edge and the blur strength.
native/runtime/ contains three prebuilt, unsigned Windows binaries:
AuraFrameBridge.exe, VCamSample.exe and VCamSampleSource.dll.
They are committed deliberately, so that cloning the repository gives you a
working virtual camera without installing Visual Studio. They are also
reproducible: the frame bridge builds from native/AuraFrameBridge/, and the
camera host builds from the vendored native/external/VCamSample/ — whose
upstream commit, license and complete Aura diff are documented in
native/external/VENDORED.md. If you would rather not trust a committed
binary, build both yourself and run npm run stage to restage them.
Being unsigned, Windows will show "Windows protected your PC" on first run of the installer. That is expected.
electron-main.js Process orchestrator. No application logic.
src/ Backend: conference, db, image, import, llm,
play, record, speech, text, web
public/ Renderer: UI, segmentation, rendering, output
services/speech/ Local Whisper HTTP service
native/ Virtual camera: frame bridge, vendored VCamSample,
staged runtime binaries
comfy/workflows/ ComfyUI graph templates
installer/ Prerequisites wizard and shipped documentation
scripts/ Build, staging and setup scripts
docs/ User-facing documentation
| Document | Contents |
|---|---|
ARCHITECTURE.md |
System architecture and execution flow |
VIRTUAL_CAMERA.md |
Native virtual-camera architecture |
TRANSLATION.md |
Universal translation subsystem |
INSTALLER.md |
Packaging and deployment |
CLAUDE.md |
Engineering conventions for this repository |
docs/Aura Documentation v0.7/ |
User guide, install guide, licensing report |
Aura is released under the MIT License — see LICENSE.
Third-party components keep their own licenses. The bundled PP-HumanSeg model is
Apache-2.0 (public/models/paddleseg/LICENSE) and the vendored VCamSample is
MIT, Copyright (c) 2024-2026 Simon Mourier
(native/external/VCamSample/LICENSE).
Note that MIT covers Aura's own source. It does not grant rights to the external
services Aura integrates with: SerpAPI is a paid service under its own terms
and is not redistributed here. A full component-by-component inventory is in
docs/Aura Documentation v0.7/Aura Third-Party Licensing Report.md.


