Copyright (c) 2026 Arshia Keshvari. Licensed under the MIT License.

Senti is a local visual assistant. The security posture is keep camera, microphone, and model traffic on the machine.

• No cloud API for frames or transcripts. The VLM talks to Ollama at VLM_BASE_URL (default http://localhost:11434).
• No recording to disk. Live frames live in a bounded in-memory buffer and a short selection buffer. They are discarded when the app exits.
• Permissions. macOS camera (always) and microphone (only if VOICE_ENABLED=true) are requested through Qt. Usage strings live in resources/Info.plist.
• No telemetry, accounts, or analytics.

If you point VLM_BASE_URL at a remote host, you are choosing to send JPEG frames there. The default is localhost.

• Silent screen or camera capture in the background
• Upload of images, audio, or transcripts to a Senti-operated service
• Persistence of conversation history across launches (memory is process-local)

Please do not open a public GitHub issue for a vulnerability that could expose camera frames, microphone audio, or a local RCE path.

Email or message the author, Arshia Keshvari, via the GitHub profile attached to TeslaNeuro/Senti, and include:

• Affected version / commit
• What an attacker would need locally
• A minimal description of the issue (not a public exploit)

We will work with you on a fix and a disclosure timeline.

Install from this repository and requirements.txt (or pyproject.toml metadata) inside a venv. Review .env before enabling OCR, TTS, or voice. YOLO, EasyOCR, and Whisper weights are downloaded by those libraries from their upstreams on first use.

Optional yolo-mlx is AGPL-3.0 and is not in the default install. Only pull it in with pip install "yolo-mlx[tracking,convert]" if you want YOLO_RUNTIME=mlx. Do not commit downloaded weights (*.pt, *.npz, *.safetensors).