Four things it does: (1) records every LLM call as a tamper-evident audit trail, (2) enforces policies in real time — rate limits, budget caps, tool restrictions, kill switches, (3) redacts PII from traces before they hit your observability backend, (4) groups traces into task-level episodes you can replay.
The demo lets you run four scenarios — normal request, runaway loop (watch the kill switch save ~$47), PII in prompt (SSN/CC redacted), and a dangerous tool call (blocked before it reaches the LLM).
Setup is git clone + docker compose up + pip install air-blackbox-sdk. One wrapper function, no changes to your agent code. Integrations for LangChain, CrewAI, OpenAI Agents SDK, and AutoGen.
22 repos, 700+ tests, Apache-2.0: https://github.com/airblackbox/gateway