A Claude Code plugin that gates every tool call before it runs, built on a Lean-verified engine that proves one property no test can give: no policy, not even one that allows everything, can let the agent disable its own guard. The engine decides allow, ask, or deny over a policy given as data; it models unseen tools and MCP calls generically and delegates containment of executed code to an OS sandbox.
Projects
A multi-agent triage system that screens feature requests for security, privacy, and GRC risk and routes each one to the right review team via GitHub Issues. It runs as a GitHub Action, dispatching three specialist Pydantic AI agents in parallel behind a deterministic orchestrator that keeps the LLM out of the control flow.
An agentic security-assessment suite for Claude Code that threat-models a real system and produces a threat model, privacy impact assessment, compliance gap analysis, and a product-grade diagram set. Designed spec-first with OpenSpec, with a reference-free evaluation harness that verifies the agent output without any per-target answer key.
An experiment that wires Microsoft Research's Memora into coding agents (Claude Code, Codex, Cursor, Omnigent) over the Model Context Protocol. Memora searches on compact cues and returns the matching detail verbatim; the bridge keeps that store and retrieval and swaps in a coding-tuned extractor, so a harness captures and recalls the reasoning worth keeping without a new service or key.