air-gate

Policy engine + human-in-the-loop tool gating + tamper-evident HMAC-SHA256 audit chain for AI agents. Helps satisfy EU AI Act Article 14 (Human Oversight) and Article 12 (Record-Keeping).

Overview

air-gate sits between your AI agent and its tools. Every tool call is checked against a policy and recorded to an append-only, cryptographically signed audit chain. Low-risk actions run automatically; risky ones (send email, delete data, execute SQL) pause and wait for a human decision before the tool runs.

Two ways to use it:

• Local mode — everything runs in-process. Zero servers. Great for a single agent, tests, or embedding the audit chain directly in your app.
• Server mode — run the Gate proxy; approvals happen in Slack (or via the HTTP API). Use this for multi-agent setups and real human-in-the-loop review.

Install

pip install air-gate              # core: GateClient + policy + audit chain
pip install "air-gate[server]"    # adds the FastAPI proxy + Slack approvals
pip install "air-gate[langchain]" # adds the LangChain tool wrapper

Quick Start (local mode)

from air_gate import GateClient

# Local mode: no server. Events are policy-checked, signed, and chained on disk.
gate = GateClient(
    signing_key="use-a-real-secret",
    storage_path="gate_events.db",
    policy_config={
        "default": "require_approval",
        "rules": [
            {"name": "search",  "action_type": "search", "decision": "auto_allow"},
            {"name": "emails",  "action_type": "email",  "decision": "require_approval"},
            {"name": "deletes", "action_type": "db_delete", "decision": "block"},
        ],
    },
)

result = gate.check(
    agent_id="recruiting-agent",
    action_type="email",
    tool_name="send_email",
    payload={"to": "jane@example.com", "subject": "Hello"},
    input_context="Agent matched Jane as a 92% fit",
)

# check() returns one of: "auto_allowed", "pending_approval", "blocked"
if result["decision"] == "auto_allowed":
    send_the_email()
elif result["decision"] == "blocked":
    print("Blocked by policy:", result["reason"])
else:  # pending_approval
    # A human approves out-of-band (Slack, API, another process), then:
    gate.approve(result["event_id"], authorized_by="alice@company.com")

# Verify the audit chain at any time
print(gate.verify())   # {"valid": True, "events_checked": N, "errors": []}

Policy configuration

Rules are evaluated in order — first match wins. If nothing matches, the default decision applies. A rule's decision is one of auto_allow, require_approval, or block. Match on any combination of agent_id, action_type, and tool_name; a field left unset matches anything.

# gate_config.yaml
policy:
  default: require_approval          # safest default: humans approve everything

  rules:
    - name: allow-read-only
      action_type: db_read
      decision: auto_allow

    - name: block-delete
      action_type: db_delete
      decision: block

    - name: approve-emails
      action_type: email
      decision: require_approval
      max_per_hour: 50               # optional rate limit (blocks over the cap)
      max_payload_size: 100000       # optional payload byte cap

gate = GateClient(config_path="gate_config.yaml")

Human-in-the-loop (blocking)

When you wrap tools with an integration, a require_approval action blocks the tool call until a human decides, then runs the tool only if approved. If it is rejected — or no decision arrives before the timeout — the wrapper fails closed and the tool never runs.

LangChain

from langchain_community.tools import DuckDuckGoSearchRun
from air_gate.integrations.langchain import GatedTool

gated_search = GatedTool(
    tool=DuckDuckGoSearchRun(),
    agent_id="research-agent",
    gate_url="http://localhost:8000",  # server mode; omit for local mode
    action_type="search",
    wait=True,        # block on pending approval (default)
    timeout=300,      # seconds to wait before failing closed
)

# Use gated_search anywhere a LangChain tool is expected.

OpenAI Agents / plain function tools

from air_gate import GateClient
from air_gate.integrations.openai_agents import gated_tool

gate = GateClient(server_url="http://localhost:8000")

@gated_tool(gate=gate, agent_id="assistant-v1", action_type="email", wait=True)
def send_email(to: str, subject: str, body: str) -> str:
    """Send an email."""
    ...  # runs only after a human approves

You can also wait manually without a wrapper:

r = gate.check("agent", "email", "send_email", payload={"to": "x@y.com"})
if r["decision"] == "pending_approval":
    outcome = gate.wait_for_decision(r["event_id"], timeout=300)  # blocks
    if outcome == "approved":
        send_the_email()

Server mode + Slack approvals

uvicorn air_gate.proxy:app --host 0.0.0.0 --port 8000
# or: docker compose up

When an action needs approval, Gate posts a message to Slack with Approve / Reject buttons; the click is recorded to the signed chain. Key endpoints:

Configuration (environment variables)

Approve/reject over HTTP with the token:

curl -X POST http://localhost:8000/actions/$ID/approve \
  -H "Authorization: Bearer $GATE_APPROVAL_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"authorized_by": "alice@company.com"}'

Audit chain & tamper-evidence

Every action and every human decision is a signed entry, chained to the one before it (previous_hash). The chain is append-only: approvals and rejections are recorded as new signed decision entries, never edits, so a resolution can never rewrite or break earlier history. Editing, deleting, or reordering any entry fails verification.

gate.verify()   # {"valid": True/False, "events_checked": N, "errors": [...]}

Note: HMAC uses a shared secret, so anyone holding GATE_SIGNING_KEY can both sign and verify. It protects against tampering by parties without the key; it is not a substitute for asymmetric signatures or an external anchor if you need to prove integrity to a party who must not hold the signing key.

CLI

air-gate demo               # self-contained demo (no server needed)
air-gate verify PATH        # re-verify an existing .db or .jsonl chain
air-gate version

PII redaction (optional)

The server can redact PII from payloads before they enter the audit chain and attach a GDPR Article 30 processing manifest. Enable with GATE_PII_REDACTION=true (default) and choose a method via GATE_PII_METHOD (hash_sha256, mask, remove, tokenise). See air_gate/pii.py for the multi-vertical detectors (recruiting, finance/PCI, healthcare/HIPAA, legal) and Article 17 erasure lookup.

Part of AIR Blackbox

air-gate is one component of the AIR Blackbox ecosystem for EU AI Act compliance:

• air-gate — tool gating + audit chain (Articles 12, 14)
• air-trust — trust layers and compliance tooling
• air-compliance — automated Article 9–15 scanning
• air-docs — model cards, consent logs, audit trails

License

Apache License 2.0. See LICENSE.

Questions? Open an issue on GitHub.