MCP servers
Connect any app, data source, or custom API to your agents and to external AI tools like Cursor and Claude, through a single governed layer.
MCP (Model Context Protocol) is how AI systems call tools. Airia sits on both sides of that standard: it connects your agents to catalogue and custom MCP servers, and it exposes those same connections to external clients like Cursor and Claude Code.
One governed layer handles approval, credentials, and which tools an AI can call, whether that AI is an agent built inside Airia or a coding assistant running on someone's laptop.
Two ways to connect
Every integration in Airia is built from the same three ingredients: an app, its credentials, and the specific tools you expose. What changes is scope and where it's used.
| Deployment | Gateway | |
|---|---|---|
| Scope | One app, for use inside Airia agents | One or more apps, bundled behind a single endpoint |
| Used by | An AI Model step inside an Airia agent | External MCP clients (Cursor, Claude Code, Claude Desktop, etc.), or an agent's tool section |
| Visibility | Scoped to a project, or all projects | Personal (just you) or Tenant (your whole org) |
| Who can create one | Any project member with access | Any user (Personal); Platform Admin or Security Admin (Tenant) |
See Deployments vs. Gateways for the full comparison, and Gateway/Deployment Creation for how to build either one.
Built-in governance
Because every connection, whether it's feeding an internal agent or an external client, passes through the same layer, a few things come for free:
- A single approval gate. Admins decide which of Airia's 1,200+ catalogue servers (plus any custom ones) are usable at all, before anyone can add them to a Gateway or Deployment. See Server Management.
- Prompt injection screening. Every tool's name and description is scanned for hidden prompt injection attempts before it's exposed. See Tool Scanning.
- Full activity visibility. Every tool call, whether it succeeded, was denied, or errored, is logged and searchable. See MCP Monitoring.
- Context-efficient scaling. Radar lets a Gateway with hundreds of tools stay lightweight by letting agents search for what they need instead of loading every tool definition up front.
- Org-specific know-how. The Instructions Tool attaches your team's own conventions, naming, and workflows to a Gateway or Deployment, so agents use your tools the way your team actually works.
What's in this section
| Group | What you'll find |
|---|---|
| Admin Controls | Approving servers, creating and editing Gateways and Deployments, credentials, tool scanning, Radar, Instructions, custom servers, and SpecLink |
| End User Usage | Connecting a Gateway to Cursor, Claude Code, or Claude, and recovering a broken credential |
| Popular MCP Servers | Setup guides for commonly connected servers like Microsoft Graph, Airtable, Snowflake, Jira, Confluence, Box, and Brave Search |
Getting started
If you're an admin:
- Approve the servers your organization is allowed to use.
- Create a Gateway or Deployment and connect credentials for the apps you need.
- Share the Gateway with your team, or attach the Deployment to an agent.
- Keep an eye on usage with MCP Monitoring.
If you're an end user:
- Connect a Gateway to Cursor, Claude Desktop, or Claude Code.
- Add a Deployment to an agent if you're building agents rather than using an external client.
- If a connection ever stops working, Credential Recovery gets you back up in a click.