irongate 0.0.2

Deterministic runtime security layer for AI agents. Enforces policies for file access, network requests, process execution, and tool invocation.

IronGate

Deterministic runtime security for Python AI agents

IronGate is a local, in-process policy enforcement layer for Python-based AI agents. It intercepts file, network, process, and tool operations at the Python API level and applies deterministic allow/review/block decisions based on a JSON policy file.

Quick Start

pip install irongate

Protect a Script

irongate run my_agent.py

All file, network, process, and tool operations will be intercepted and enforced against your policy.

Use as a Library

from agentshield import shield

with shield.protect():
    agent.run(query)

Features

• ✅ Deterministic enforcement - No AI/ML heuristics, pure policy-based decisions
• ✅ Multiple enforcement layers - Audit hooks + monkeypatching for defense in depth
• ✅ Local-only runtime - All processing happens in-process, no external services
• ✅ Fine-grained control - Policy rules for file, network, process, and tool operations
• ✅ Human-in-the-loop - Dashboard for reviewing and approving operations
• ✅ Audit trail - Hash-chained SQLite logs for tamper evidence
• ✅ Deny-by-default - Unknown operations are blocked by default

Intercepted Operations

• File: open(), read(), write(), unlink(), rename(), mkdir(), and more
• Network: requests, httpx, urllib, aiohttp, socket, ssl
• Process: subprocess, os.system, os.exec*, os.spawn*
• Tools: Custom function invocations through LangChain or generic wrappers

Installation

pip install irongate

Requires Python 3.8+ (audit hooks introduced in Python 3.8).

Documentation

• Full documentation: https://github.com/pratikacharya1234/Irongate#readme
• Security model: https://github.com/pratikacharya1234/Irongate/blob/main/SECURITY_MODEL.md
• Threat model: https://github.com/pratikacharya1234/Irongate/blob/main/THREAT_MODEL.md

License

MIT License - see LICENSE for details.

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

---

Status: Experimental (v0.1.0) for evaluation and feedback. Not recommended for production use without thorough testing.