MakerChecker: An Open Source Security Layer for AI Agents

MakerChecker is an open-source security layer that sits in front of AI agents built on frameworks like LangChain, Claude SDK, or CrewAI, acting as a checkpoint before tool calls and as a cryptographically signed audit ledger afterward. Its `mc scan` command inspects agent code to detect consequential actions—data deletion, money movement, shell execution, secret exfiltration—classifies them by risk against real-world incidents, and can auto-generate governance code.

The system enforces a deny-by-default model: agents act only through defined roles with explicitly granted skills, and high-risk actions are routed to separate roles (e.g., a risk desk or medical reviewer), structurally preventing an agent from approving its own work. An optional self-hosted server adds a Postgres-backed approval inbox, review console, and a hash-chained, Ed25519-signed audit trail—any tampering breaks verification, and exported bundles can be verified offline without trusting the originating system.

The project splits into three independently adoptable packages—the scanner, an embeddable enforcement library, and the full server—all sharing the same signed audit format. Ready-made connectors for LangChain and Claude Agent SDK, plus runnable examples in pharmacovigilance, medical-device complaint triage, and cash reconciliation, demonstrate how engineers can enforce hard authorization boundaries and meet auditability requirements when deploying agentic automation in production.