Sentinel

Sentinel is the missing security layer for AI agents: DevTools that intercept every tool call before it executes. The problem is simple: prompt injection hidden inside an email, ticket, or document can hijack an autonomous agent into sending money, leaking PII, or sending sensitive data to an attacker-controlled domain. Today, most teams only discover this after the fact by reading logs. Sentinel gives them a real-time debugger and firewall for agent actions. Sentinel uses two layers of defense. A deterministic Policy DSL decides allow, pause, or block in under 5 milliseconds. For gray-zone actions, Pre-cog uses Opus 4.7 with extended thinking to reason through the causal chain before anything irreversible happens. When an incident happens, teams can replay the timeline, inspect what was blocked, compare the blast radius, fork alternate outcomes, and synthesize new policies from bypasses. The Adversarial Evolution Arena lets Opus instances attack and defend live, mutating attacks, generating defenses, and raising the system’s Trust Score from D to A+ in minutes. Built solo in five days. TypeScript end-to-end, Next.js 16 + React 19, Hono + SQLite, Anthropic SDK with streaming extended thinking, MCP server for Claude Desktop, deployed on Fly.io and Vercel.