Watch MAREF in action.

Full product introduction — whiteboard hook, risk alert, IDE showcase, brand close. 100 seconds of governance storytelling.

21-second speed intro: what MAREF does and why it matters for agent governance.

A 97-second animation unpacking the multi-agent recursive evolution framework: identity, permissions, audit, to self-evolving defense — layer by layer, how the governance system contains runaway.

Meta AI researcher's agent deleted hundreds of emails despite 3 "STOP" commands. Real story.

CURL shut down its bug bounty program — 95% of submissions were AI-generated spam.

470 PRs scanned: 45% of AI code had vulnerabilities. Severity 1.7× higher than human code.

88% of enterprises reported AI agent incidents last year. MAREF gives every agent a constitution.

3 consecutive anomalies trigger auto-circuit-break + 30s cooldown. Identity → Permission → Audit in 0.3s.

Every operation signed, traceable, immutable. Audit comes? Export the report. Done.

Input sanitization → tool audit → permissions → sandbox → safety gate → policy tree → threat detection → telemetry.

Gray Code FSM: 10 states, Hamming distance = 1, HALT is absorbing. Model-checked with TLA+.

Red-blue adversarial evolution with Lyapunov-style convergence monitoring. TLA+ model-checked.

SM2/SM3/SM4 national cryptography built-in. Every operation signed. Export and done.

Governance sits in the middle. Works with LangGraph, CrewAI, AutoGen, Claude Code, Cursor, Trae over MCP + A2A.

TLA+ model checking, defense-in-depth, zero trust, Gray Code FSM, self-evolution, national crypto, observability.

docker run maref/lite — 12 preset governance rules, zero config. Auto-block dangerous operations.

pip install maref. Open source, Apache 2.0. 11,000+ tests, production ready. Make失控可控.