Doberman

Framework-agnostic middleware for approving or denying AI agent tool calls before they execute.

Doberman sits between an AI agent and its tools. Instead of allowing an agent to directly execute a function, the call is first routed through Doberman, which decides whether to allow or block it based on user-defined policies or an interactive approval prompt.

Features

• Approve or deny tool calls before execution
• Allow once / deny once
• Always allow / always deny
• Persistent permission storage (policies.json)
• CLI for managing permissions
• Framework-agnostic design
• Zero external runtime dependencies

Installation

pip install doberman-agent

Or install from source:

git clone https://github.com/DakshSharma304/Doberman.git
cd Doberman
pip install -e .

Quick Example

from doberman import doberman

tool_call = {
    "tool": "gmail",
    "args": {
        "action": "send_email",
        "to": "alice@example.com",
        "subject": "Hello",
        "body": "Hi!"
    }
}

tools = {
    "gmail": gmail_tool
}

result = doberman(tool_call, tools)

If the permission has never been seen before, Doberman asks the user:

Doberman permission request

Tool: gmail
Action: send_email

1 = Allow once
2 = Always allow
3 = Deny once
4 = Always deny

Future requests follow the stored policy automatically.

Example

A complete Gemini + Gmail example is included:

examples/gemini_gmail_agent.py

Install the example dependencies:

pip install -r requirements-examples.txt

CLI

List permissions:

doberman list

Allow permissions:

doberman allow gmail.send_email

Allow multiple permissions:

doberman allow gmail.send_email gmail.read_latest

Deny permissions:

doberman deny gmail.send_email

Remove permissions:

doberman remove gmail.send_email

Project Structure

doberman/
    __init__.py
    cli.py
    doberman.py
    policy.py

examples/
    gemini_gmail_agent.py

Roadmap

• Desktop approval dialog
• Wildcard policies (gmail.*)
• Conditional policies
• Framework adapters (OpenAI, Anthropic, LangGraph, CrewAI)
• Logging and auditing
• PyPI release

Contributing

Issues and pull requests are welcome.

License

Licensed under the GNU GPL v3.