All work
AI Agents & MCP

Covenant

A governance and immunity layer for tool-using AI agents.

Compiles plain-English rules into typed policy, intercepts every tool call deterministically, and emits hash-linked audit receipts.

What makes it serious
  • Compiles plain-English rules into typed, enforceable policy
  • Deterministically intercepts every agent tool call
  • Emits hash-linked, tamper-evident action receipts
  • Turns human corrections into regression evals and approval-gated policy patches

Covenant is a governance-and-immunity layer for tool-using AI agents. It compiles plain-English rules into a typed policy, then deterministically intercepts every tool call an agent tries to make, allowing, blocking, or gating it. Each action produces a hash-linked receipt, so there is a tamper-evident audit trail of exactly what the agent did and why.

It also learns: when a human corrects the agent, Covenant turns that correction into a regression eval plus an approval-gated policy patch, so the same mistake is caught next time. It was built for OpenAI Build Week 2026 in the Dev Tools track, with a live demo deployed on Vercel. The stack is Next.js 15, TypeScript, Zod, and the OpenAI Responses API, tested with Vitest and Playwright.

Stack
Next.js 15TypeScriptZodOpenAI Responses APIVitestPlaywrightVercel
#ai-governance#agents#policy-engine#audit#guardrails#openai