Hey Product Hunt! 👋 I built AgentSentinel because as autonomous coding agents (Claude Code, Cursor, Devin) and Model Context Protocol (MCP) tools gain full execution access to our terminals, filesystems, and production databases, prompt injection can lead to catastrophic damage.

Existing tools rely on fuzzy LLM judges that can be jailbroken. AgentSentinel is fundamentally different: it sits directly in the tool execution lifecycle as a deterministic, AST-level zero-trust circuit breaker.

🛡️ Key Highlights:

• Sub-2ms Latency: Zero perception lag, evaluated locally in-memory with zero cloud telemetry.

• Exploit Interceptor: Automatically catches reverse shells (curl|bash), fork bombs, and raw disk writes (dd).

• Database Circuit Breaker: Intercepts unconstrained DROP TABLE, TRUNCATE, and mass DELETE without WHERE clauses.

• Sandboxed Workspace Jails: Strict filesystem boundaries preventing directory traversal (../../) and scrubbing API keys/bearer tokens.

• Universal Compatibility: Plug-and-play middleware for MCP Servers, Cursor, Claude Desktop, LangChain, and Vercel AI SDK.

📊 Battle-Tested & Research Backed:

• 1,919 automated unit tests passing across 536 test suites.

• 691 merged security modules & Apache-2.0 open-source.

• ORCID Researcher Backed: https://orcid.org/0009-0008-2818-6579 (Formal CITATION.cff & Quant White-Paper included).

Try the live terminal exploit simulation right now:

npx @sparkysolutions/agentsentinel demo

Check out the GitHub repo: https://github.com/sparkysolutions10-pixel/agentsentinel

Would love your feedback, bug reports, and roadmap ideas! 🚀