Airlock is a secure gateway that connects AI agents to your APIs. It provides:
- Pre-built Integrations: Connect to GitHub, Linear, Notion, Google Calendar, and 190+ other services instantly. See Pre-built Integrations for the full list.
- Custom API Support: Bring any API with an OpenAPI specification
- MCP Protocol Support: Native integration with Claude Desktop, Claude Code, Augment, and other MCP clients
- Approval Workflows: Human-in-the-loop controls for sensitive operations, with notifications by email, Slack, or Microsoft Teams
- AI Use Policy: Publish a company AI use policy and require members to sign it before their agents can run tools
- Per-User Authentication: Each user connects with their own credentials
- Policy Engine: Fine-grained access control for each API operation
- Security Monitoring: Automated anomaly detection and auto-blocking
How It Works
AI Agent ──── MCP Protocol ────→ Airlock ────→ Your API / Service
│
├── Policy Evaluation
├── Approval Workflow (if required)
├── Security Monitoring
└── Audit Logging
- Create an Integration: Pick a pre-built integration or upload an OpenAPI specification
- Set Policies: Define which operations are allowed or require approval
- Connect Users: Each user authenticates via MCP OAuth and connects their own credentials
- AI Calls API: The AI agent calls tools, airlock evaluates policies and routes accordingly
Key Concepts
Integrations
An integration in airlock represents an API or service you want to expose to AI agents. Airlock supports several integration types:
| Type | Description |
|---|---|
| Pre-built Integration | Connect to services like GitHub, Linear, Notion, etc. with minimal configuration |
| Custom OpenAPI | Bring any REST API with an OpenAPI specification |
| MCP Proxy | Wrap an MCP server you already run behind airlock's policies, approvals, and audit trail |
| Airlock Management | Auto-created in every organization — lets agents inspect and manage your airlock integrations, usage, and configuration |
| airlock memory | Per-organization knowledge graph for entities and relationships, queried via the memory_* tool family |
| airlock code graph | Index your GitHub repositories and let agents query architecture, call graphs, and cross-repo references |
Policies
Policies determine what happens when an AI agent tries to call an API operation:
- ALLOW: The operation executes immediately
- DENY: The operation is blocked outright and never executes
- REQUIRE_APPROVAL: The operation waits for human approval before executing
When you create an integration, its tools start as ALLOW — with one exception: the built-in airlock memory and airlock code graph integrations pre-set their write tools to REQUIRE_APPROVAL. You can change any tool from the integration's Tools & Policy tab.
Tools that appear later, when an upstream service adds them, are handled by your organization's new-tool policy: by default they are visible but every call requires approval until an admin marks them reviewed. Admins can switch this to auto-enable or auto-hide under Settings → Security.
AI Use Policy
Admins can publish a company AI use policy that members must sign. In enforce mode, a member who hasn't signed the current version cannot execute tools through airlock until they do. See AI Use Policy.
Approval Workflows
When a policy requires approval:
- Airlock creates a pending request
- Designated approvers are notified — always by email (unless an admin turns it off), and additionally in Slack or Microsoft Teams if those channels are connected under Settings → Organization → Notifications
- Approvers can approve or reject from the web UI
- The AI agent receives the result and can proceed or handle the rejection
Organization-Wide Access
Airlock provides a single organization-wide MCP endpoint that lets agents discover and use tools across all your integrations. Agents use meta-tools (list_services, search_tools, describe_tools, execute_tool) to navigate available APIs dynamically, skill tools (activate_skill, read_skill_attachment) to load skills, and agent tools (list_agents, describe_agent, export_agent) to discover your organization's published agents.
Security
Airlock monitors API usage for anomalous patterns and can automatically block suspicious activity. Security blocks are visible in the Control Room and can be managed by admins.
API Endpoints
Airlock exposes a single MCP HTTP endpoint per organization at mcp.air-lock.ai. Both POST / and POST /org/{slug} are served by the same handler — most MCP clients work with either URL. See the MCP Protocol reference for details.
Next Steps
- Quickstart - Get up and running in 5 minutes