MakerChecker: An Open-Source Security Layer for AI Agents

MakerChecker is an open-source security layer designed to bring governance to AI agents. It sits between an agent and its tool calls, enforcing deny-by-default permissions, requiring human approvals for high-risk actions, and maintaining a tamper-evident audit trail. The project provides a static scanner to identify risky agent capabilities, an embedded library to wrap tools with governance, and an optional self-hosted server for centralized enforcement and review.

How It Works

MakerChecker is structured around three independent packages. The static scanner (mc scan) analyzes agent code to classify actions by risk level, flagging operations like data deletion, money movement, or shell command execution. The embedded library (@makerchecker/embedded) lets developers define roles, skills, and agents, then wrap tool calls so that only granted actions execute. The self-hosted server provides a human approval inbox, a review console, and a verifiable audit log.

Key Features

• Deny-by-default enforcement: Agents can only execute skills explicitly granted to their role. Unauthorized calls are rejected before execution.
• Human-in-the-loop approvals: High-risk actions require separate human sign-off, preventing agents from approving their own work.
• Cryptographically signed audit logs: Every decision and tool call is recorded in a hash-chained, Ed25519-signed log. Tampering breaks verification, and bundles can be verified offline.
• Framework integrations: Drop-in connectors for LangChain, Claude Agent SDK, and TypeScript/Python SDKs make it easy to add governance to existing agent workflows.

Use Cases

The project includes several example implementations, such as pharmacovigilance case processing, medical-device complaint triage, oncology patient access, and daily cash reconciliation. These demonstrate how MakerChecker can enforce segregation of duties and regulatory compliance in real-world agent deployments.

Getting Started

To scan an existing agent codebase for risk, run npx @makerchecker/scan .. To embed governance into a new agent, install @makerchecker/embedded and define roles and skills. For centralized enforcement with a human approval inbox, run docker compose up to start the self-hosted server. The project is available under a dual license (Apache 2.0 for non-production use, commercial for production).