Under construction - Sentinel is still in development

Sentinel

Status - In development. ⛏️

A secure, self-evolving personal AI agent with zero external daemons, installed via a single .exe.

Sentinel is an autonomous AI agent that runs as a native Windows desktop application. It wraps the API of any LLM in an OpenAI-compatible API, including both cloud and local models. With optional local LLM fallback in the event that a cloud model become unavailable during operation for any reason. Sentinel offers a branded Open WebUI chat interface in a native pywebview window. The agent reasons, uses tools, maintains memory, evolves its identity over time, and communicates across multiple channels.

✨ Features

Capability	Description
Agentic loop	Trigger → prompt assembly → LLM call → tool dispatch → loop, with hard iteration, loop, and tool timeouts.
Three-tier memory	Working, episodic, and semantic memory over SQLite with `sqlite-vec` vector search.
Private by default	Embeddings are generated locally on-device — memories are never sent to a third-party service.
Identity evolution	The agent's self-description is revised over time from its own accumulated experience.
Skill framework	Sandboxed tool and channel skills, each declared by a `skill.json` manifest, with a CLI for scaffolding, validation, and installation.
Multi-channel	Chat, email, Telegram, WhatsApp, and phone — all normalised through a single channel router.
Voice	Local-first speech-to-text (`faster-whisper`) and text-to-speech (Piper), with cloud fallback.
Security	OS-level sandboxing via Job Objects and ACLs, OS keyring credential storage, encryption at rest, and a domain-allowlisted egress proxy.
Single-file install	Embedded Python, a PyInstaller launcher, and an Inno Setup installer. No external daemons, no service, no administrator rights.

🧩 Architecture

Sentinel runs entirely in-process or as managed subprocesses — there is no external daemon, message broker, or database server to install. It is a desktop application running in your own user session, not a background Windows service.

Layer	Technology
Language	Python 3.12 (async throughout)
API gateway	FastAPI + uvicorn, exposing an OpenAI-compatible endpoint
Language model	Anthropic Claude API (primary), Ollama (optional local fallback)
Embeddings	`all-MiniLM-L6-v2` via ONNX Runtime — local, offline, 384-dimensional
Storage	SQLite in WAL mode, `sqlite-vec` for vector search, `diskcache` for caching
Interface	Open WebUI in a native `pywebview` window, with a `pystray` system tray
Isolation	Windows Job Objects, ACLs, and an egress proxy with a domain allowlist

✅ Requirements

Windows 10 or 11 (64-bit)
Python 3.12 (bundled with the installer build; required only for running from source)
An Anthropic API key
Optionally, a local Ollama installation for offline LLM fallback

No Docker, no WSL, no Node.js, and no system-wide Python installation are required.

🚀 Getting Started

# Install dependencies
pip install -r requirements.txt

# Run the application
python -m sentinel start

Credentials are never read from the repository or from configuration files — Sentinel stores them in the operating system keyring on first run.

💻 Development

# Run the automated test suite
pytest -m "automated" --tb=short -q

# Type checking
mypy src/sentinel/ --strict

# Linting
ruff check src/ tests/

🙏 Acknowledgements

Open WebUI (GitHub)
- Sentinel's chat interface is Open WebUI, by Timothy Jaeryang Baek and contributors.
- Sentinel runs Open WebUI as a managed subprocess inside a native window and points it at the agent.
LLM APIs
- Anthropic APIs.
- Ollama model APIs.
SQLite
- with sqlite-vec
FastAPI
Pydantic
pywebview

📄 License

Sentinel bundles no Open WebUI source. It is installed as an ordinary dependency and remains under its own licence, whose branding terms Sentinel observes.