Trust layer for AI

Ship AI features with policy, redaction, approval, and audit — from one gateway.

Airlock governs AI requests and the actions agents intend to take — enforcing policy, scope boundaries, human approval, and auditability before execution.

Integrate in under an hour. One dependency, one env var, one endpoint.

Bring your own OpenAI key — connection guide. Airlock governs every request; your provider bills you directly.

Animation of a request passing through Airlock's governed pipeline: policy, redaction, approval, audit. — Every request moves through the same governed path.

Every request moves through the same governed path.

Principles

Airlock makes trust boundaries explicit instead of implicit.

Policy first

Every request — and every action an agent intends to take — is checked against a versioned policy and its scope. No match, no call.

Sensitive data aware

Built-in detectors for email, phone, SSN, credit card, and custom regex. Redact, warn, block, or route for approval — per policy.

Human in the loop

Send high-risk requests and elevated agent actions to a human approval queue. Approvals resume the pipeline transparently — no bespoke plumbing per app.

Immutable audit

Every request and evaluated action writes a tamper-evident audit event with correlation and agent run IDs, the policy version, and the decision reasoning.

What counts as an application?

In Airlock, an application is any product, service, workflow, or integration that sends AI requests through Airlock.

That can include internally built software, backend services, customer portals, support tools, document workflows, AI agents that call tools, or integrations connected to enterprise systems like CRMs, ticketing platforms, or knowledge bases.

The key requirement is control of the request path. If your team can route the AI call through Airlock — directly, through an SDK, or through middleware — Airlock can govern it. Agents can also describe the action they intend to take — the tool, resource, and operation — so Airlock can allow it, block it, or require human approval first. The agent runtime still performs the action; Airlock never executes tools on its behalf.

Airlock is not designed to inspect closed SaaS tools where the AI request is made entirely inside a vendor's system and cannot be routed through your gateway.

Internally built support portal

Backend document summarization service

CRM or ticketing workflow via middleware

Customer-facing AI feature

Internal policy assistant

Integrate in under an hour

Point your existing OpenAI-compatible client at the Airlock gateway. Attach your API key. Every call is governed, audited, and returned transparently.

# One HTTP call, fully governed.
curl -X POST "$AIRLOCK_URL/api/public/v1/completions" \
  -H "authorization: Bearer $AIRLOCK_API_KEY" \
  -H "x-correlation-id: job-4821" \
  -H "content-type: application/json" \
  -d '{
    "purpose": "support_reply",
    "model": "gpt-4o-mini",
    "messages": [{"role": "user", "content": "..."}]
  }'