Doberman

Doberman is an open-source runtime authorisation layer for AI coding agents, such as Claude Code, Codex and Cursor. Coding agents can carry out harmful actions, such as deleting files, leaking credentials or following hidden instructions planted in content they read (prompt injection). Doberman addresses these risks by checking every action an agent takes before it runs. Unlike guardrails that only inspect prompts, it sits directly between the agent and its tools, so a blocked action never executes. It is currently in an early (beta) stage of release.

Doberman connects to agents through a proxy for the Model Context Protocol (MCP), the standard interface between agents and their tools, or through hooks built into the agent's host. Each tool call receives one of three verdicts. Routine actions pass straight through. Sensitive actions are paused for human approval. Dangerous actions are blocked. An optional check also screens user input for known prompt-injection patterns before the model processes it.

Two design principles guide the tool. It fails closed, meaning any error or uncertainty results in the action being denied. It is also "raise-only": protections can tighten automatically, but any weakening requires explicit human approval, protected by a password or two-factor authentication, and is recorded in an audit log. Teams can commit policies to their code repository and review them like code. Users can adjust strictness levels and add custom rules

Auto-discovered on 2026-09-23 by OECD Catalogue Automation