Doberman-Core

Doberman is an open-source runtime security layer for AI agents that monitors tool calls, enforces execution policies and stops dangerous actions before they reach your system. It sits between an agent and its tools, assigning every action a PASS, AUTH or BLOCK verdict before execution.

Doberman combines deterministic guardrails with adaptive risk detection to protect against destructive commands, prompt injection, secret exposure, unsafe file operations and unauthorized tool use. It supports AI coding agents, MCP clients and autonomous workflows while preserving useful agent autonomy instead of forcing human approval for every action.

Built for developers and security teams, Doberman provides runtime policy enforcement, step-up authorization, audit logs and agent activity monitoring across tools like Claude Code, Codex, Cursor and other MCP-compatible systems. The open-source core is local-first and designed to act as a final security boundary when model-level guardrails fail.