AIR Controls

April 3, 2026 · View on GitHub

See what your AI agents actually do.

Runtime visibility for AI agents: action timeline, intent translation, anomaly detection, and guardrails. Part of the AIR Blackbox ecosystem.

pip install air-controls

Why

Companies are deploying AI agents that send emails, update CRMs, and run workflows. Nobody knows what these agents are actually doing. Logs exist but are useless. Observability tools are too technical. There's no "understandable layer."

AIR Controls fixes that.

Quick Start

LangChain / LangGraph

from air_controls import ControlsCallback

cb = ControlsCallback(agent_name="sales-bot")
chain.invoke({"input": "..."}, config={"callbacks": [cb]})
# That's it. Run `air-controls status` to see the dashboard.

Custom Agents (OpenAI / Anthropic API)

from air_controls import monitor

@monitor(agent_name="my-bot")
def process_customer(query):
    response = openai.chat.completions.create(...)
    return response

# Or as a context manager with manual logging:
with monitor(agent_name="my-bot") as m:
    response = openai.chat.completions.create(...)
    m.log("api_call", "POST /v1/chat/completions", "Generated AI response")

CrewAI

from air_controls import CrewMonitor

mon = CrewMonitor(agent_name="research-crew")
result = mon.run(crew)

AutoGen

from air_controls import AutoGenMonitor

mon = AutoGenMonitor(agent_name="coding-assistant")
mon.attach(agent)

CLI

air-controls status              # Show all agents and recent activity
air-controls events sales-bot    # Show event timeline for an agent
air-controls stats sales-bot     # Detailed statistics
air-controls pause sales-bot     # Kill switch — pause an agent
air-controls resume sales-bot    # Resume a paused agent
air-controls verify              # Verify audit chain integrity

What Gets Tracked

Every agent action is logged as a structured event:

Action type: LLM call, tool use, API call, decision, error
Human summary: Plain English translation of what happened
Cost: Token usage and dollar cost per action
Duration: How long each action took
Risk score: Low / medium / high based on action type
Audit chain: HMAC-SHA256 tamper-evident chain (shared with AIR Blackbox)

Features

Feature	Status
Action timeline	✅ Shipped
LangChain callback	✅ Shipped
CrewAI monitor	✅ Shipped
AutoGen monitor	✅ Shipped
Custom agent decorator	✅ Shipped
CLI dashboard	✅ Shipped
Kill switch	✅ Shipped
HMAC-SHA256 audit chain	✅ Shipped
Web dashboard	🔜 Coming
Intent translation	🔜 Coming
Anomaly detection	🔜 Coming
Guardrails & rate limits	🔜 Coming

How It Fits

AIR Blackbox (pre-deploy)  →  Scans code for EU AI Act compliance
AIR Controls (runtime)     →  Monitors what agents actually do
                ↓
    Same HMAC-SHA256 audit chain

Local-First

Your agent telemetry never leaves your machine. Everything runs locally on SQLite. No cloud. No phone-home. No API keys needed. Set a custom DB path:

export AIR_CONTROLS_DB=/path/to/events.db

License

Apache 2.0