v3.0.0 - First public release: Sentinel AI Stack

A batteries-included personal AI butler for your home, powered by Claude and local LLMs - controlled entirely from Telegram.

This is the first public release of Sentinel AI Stack. The project has been running privately for ~2 weeks of intense development; this v3.0.0 captures the stable feature set ready for self-hosting.

What you get

Two Telegram bots

• @YourSentinelBot - the AI assistant (you talk to it; it talks to your tools)
• @YourWatchdogBot - the management plane (restart commands, alerts, status)

MetaMCP gateway aggregating 10+ tools

• Memory MCP (long-term recall across sessions)
• Reminders MCP (one-shot, recurring, multi-recipient)
• Google Workspace (Gmail, Drive, Calendar)
• OneDrive (with Azure DocIntel for PDF parsing)
• yt-dlp / gallery-dl (YouTube, Instagram, TikTok downloads)
• Google Maps (directions, place search)
• GitHub (repos, issues, PRs)
• LibreTranslate (offline translation)
• Playwright (full browser automation)

Local model inference

• LM Studio with Qwen3.6 27B (chat) + Qwen2.5-Coder-32B (coding)
• 3-way prompt classifier picks the right model per request
• Inference bridge with fallback chain when target model isnt loaded

Mini App (Telegram WebApp)

• TOTP-gated dashboard at your own Cloudflare Tunnel subdomain
• Live agent browser view (Perplexity-style screenshot streaming + history scrubber)
• Reminders manager with multi-select contacts
• Memory browser
• Pending pairings (one-tap approval for new users)
• Updates panel (tracks all upstream component versions)
• OpenClaw config + doctor panels
• Per-tester daily message caps

Watchdog Telegram bot

• Auto-restart for Docker containers, OpenClaw, LM Studio, bridges, Playwright
• CRITICAL escalation when auto-recovery fails
• DNS health, GitHub sync, daily digest

Power-conflict protection

• Bridge blocks new inference + aborts in-flight streams when a Steam game starts
• Designed for shared 650W-PSU rigs that cant safely run gaming + LLM at peak

Real credential handling

• Secrets in Windows Credential Manager (not .env files)
• Sync script regenerates .env.local from WCM at boot
• Pre-commit scanner blocks plaintext secret leaks
• Apache 2.0 with explicit patent grant

Documentation

Doc	Contents
docs/miniapp.md	Dashboard screens, auth flow, skill credentials, API reference
docs/watchdog.md	Bot commands, monitoring logic, auto-restart map, config
docs/reminders.md	APScheduler architecture, decoupled firing, multi-recipient
docs/inference-bridge.md	3-way model routing, gaming/inference power lock, availability fallback
docs/config.md	Config files, secrets management, version system
docs/llm-prompt.md	Full installation walkthrough written for an LLM to execute

Install

See the First-time setup section in the README. Requires Windows 11 + WSL2 + Docker Desktop + Python 3.11+ + LM Studio + Cloudflare Tunnel.

Reproducible "minimal demo mode" (no cloud APIs, just Telegram + LM Studio + Reminders + Memory) is on the roadmap for v3.1.

Contributing

Apache 2.0 license. Especially welcome: documentation fixes, Linux host support (community-maintained beyond this), new MCP servers, anything that reduces "first run" friction.

Open a Discussion before non-trivial PRs.

What''s next

V3 Phase 1.3 (browser co-pilot mode - clicks + typing in canvas forwarded to Playwright)
V3 Phase 2 (voice in/out, smart home control, embedded Maps panel, scheduled tool-call automation)
SG Deals MCP (Shopee/Lazada/Amazon SG search + price tracking)
V5 multi-tenant (per-tenant secrets in their own cloud, BYO-model)

Acknowledgements

Built on top of MetaMCP, OpenClaw, Playwright, LM Studio, and many open-source MCP servers - see NOTICE for full attribution.