ICEBOX

ICEBOX protects your AI agent from doing stupid things.

Install

pip install icebox-sdk
icebox init

icebox init asks what you want to protect and sets everything up. No YAML, no policy DSL, nothing to configure.

Use it

from icebox import govern

if govern("Delete EC2 Instance", target="Production AWS"):
    delete_ec2()

That's the whole SDK. Describe the action in plain words; ICEBOX decides whether it's allowed. If govern(...) is truthy, do it. If not, ICEBOX already stopped it and saved the audit entry.

When an action is held or blocked, ICEBOX tells you:

ICEBOX protected Production AWS.

Action:
  Delete EC2 Instance

Decision:
  Approval Required

Your AI agent attempted a dangerous action and was stopped.

Check your protection

icebox doctor

ICEBOX Status

✓ Daemon running
✓ Policy loaded (6 rules)
✓ Audit enabled
✓ Sandbox enabled
✓ Production AWS profile loaded

You're protected.

Examples

ICEBOX works in front of any agent framework. Recipes, not integrations:

• Claude Code
• OpenAI Agents SDK
• CrewAI
• AutoGen

Each is a thin wrapper: intercept the tool call, ask govern(), run only when allowed.

The two ideas

1. Protect something. icebox init sets up the guardrails.
2. Govern dangerous actions. govern() answers one question: is this allowed?

Everything else — sandboxes, audit trails, approval workflows — is ICEBOX's job, not yours.