SentryFlow v1.0.0 is out
The AI proposes. Policy decides what runs.
SentryFlow is a permission-aware execution layer for AI agents, built as a self-hosted MCP server (spec 2025-11-25, Streamable HTTP) with a simulated Alexa+ household.
What's in v1.0.0
- Local AI planning with Ollama (qwen2.5:3b), with a transparent rule-based fallback
- Deterministic policy engine: ALLOW / REQUIRE APPROVAL / DENY, default deny
- Owner approvals bound to a SHA-256 hash of the exact arguments: single-use and expiring
- Re-validation right before execution, so nothing runs twice
- Hash-chained audit log with a receipt for every request
- 110 automated tests, most of them adversarial (prompt injection, approval replay, argument tampering)
Project page and demo video: https://ethical0101.github.io/SentryFlow/ Source (MIT): https://github.com/ethical0101/SentryFlow
All devices are simulated. Nothing controls real Alexa or Ring hardware.
Feedback welcome in the comments!
Log in or sign up for Devpost to join the conversation.