Repolode
Submit
DobermanCore

DobermanCore/Doberman-Core

A runtime guard for AI coding agents that gates every tool call and returns PASS, AUTH or BLOCK before the action executes.

View on GitHubWebsite
Stars
247
Forks
44
Open issues
54
Last commit
9 days ago

What is it?

Doberman is an authorization and guardrail layer that sits on the execution path between an AI agent and its tools such as files, shell, MCP servers and APIs. Each tool call receives exactly one verdict before it runs: PASS for routine actions, AUTH for sensitive actions that need approval, and BLOCK for dangerous actions that never execute. It works through a transparent MCP proxy or through host hooks for Claude Code, Codex, OpenClaw and Cursor.

The design fails closed, so errors or unhandled cases deny the action, and approval prompts that get no answer resolve to a denial after a deadline. Adaptive learning can tighten policy automatically but cannot silently loosen it; weakening a policy requires an audited human approval gated by a possession factor such as TOTP or a local password. Policy can also be committed to the repo as a doberman.policy.yaml file, so changes go through code review.

The project is Python-based, installs with pip as doberman-core, and includes a CLI with setup, doctor, demo and dashboard commands, along with a benchmark harness that measures attack bypass and benign over-block rates.

Key features

  • 01Gates every tool call with a PASS, AUTH or BLOCK verdict before execution.
  • 02Works as a transparent MCP proxy in front of any MCP-compatible server.
  • 03Provides host hooks for Claude Code, Codex and Cursor (experimental).
  • 04Fails closed and never silently loosens policy without audited human approval.
  • 05Supports repo-committed policy files reviewed alongside code.
  • 06Includes a dashboard for live decisions and pending approvals.

Use cases

  • Preventing an AI coding agent from running destructive shell commands like rm -rf on a repository.
  • Stopping agents from reading .env files or leaking API keys through tool calls.
  • Requiring human approval for sensitive actions in unattended agent workflows.
  • Auditing and enforcing tool-use policy for teams deploying agents across a fleet.

Quick start

  1. 1

    Install Doberman from PyPI

    pip install doberman-core
  2. 2

    Enable shell tab completion (optional)

    doberman --install-completion
  3. 3

    Set up Doberman for Claude Code (asks which agents to guard and tunes guardrails)

    doberman setup
  4. 4

    Confirm the installation is working

    doberman doctor

Things to know

  • Status is alpha, so expect breaking changes.
  • Codex and Cursor integrations are marked experimental.
  • Approval prompts auto-deny after a timeout when unanswered.

In collections

Open-source AI agent frameworks & appsMCP serversSecurity scanners, audits & threat intel tools

Topics

#agent-security#ai-agents#guardrails#mcp#llm-security#authorization#python#ai-safety#llm#security

FAQ

What is Doberman-Core?
A runtime guard for AI coding agents that gates every tool call and returns PASS, AUTH or BLOCK before the action executes. Doberman is an authorization and guardrail layer that sits on the execution path between an AI agent and its tools such as files, shell, MCP servers and APIs. Each tool call receives exactly one verdict before it runs: PASS for routine actions, AUTH for sensitive actions that need approval, and BLOCK for dangerous actions that never execute. It works through a transparent MCP proxy or through host hooks for Claude Code, Codex, OpenClaw and Cursor.
Is Doberman-Core free to use? What license does it have?
Doberman-Core is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is Doberman-Core still maintained?
Actively maintained. The last commit to Doberman-Core was on Oct 29, 2026.
What can I use Doberman-Core for?
Preventing an AI coding agent from running destructive shell commands like rm -rf on a repository. Stopping agents from reading .env files or leaking API keys through tool calls. Requiring human approval for sensitive actions in unattended agent workflows. Auditing and enforcing tool-use policy for teams deploying agents across a fleet.
How do I install Doberman-Core?
Install Doberman from PyPI: pip install doberman-core Enable shell tab completion (optional): doberman --install-completion Set up Doberman for Claude Code (asks which agents to guard and tunes guardrails): doberman setup Confirm the installation is working: doberman doctor
Where can I get Doberman-Core?
The source code and installation instructions are on GitHub: https://github.com/DobermanCore/Doberman-Core

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects