MCP Servers Overview
MCP Servers are deployed endpoints that expose your ToolSet's tools to AI agents via the Model Context Protocol (MCP). When you deploy a ToolSet as an MCP server, Nexla provisions a dedicated endpoint that AI agents like Claude, Cursor, and ChatGPT can connect to, discover your tools, and use them to interact with your data.
What Are MCP Servers?
An MCP Server is a running instance of a ToolSet, referred to as an "export" in the Nexla Platform. Each export gets a unique URL and server key that MCP-compatible clients use to connect. Once connected, the client can list all available tools in the ToolSet and invoke them as part of AI agent workflows.
MCP Servers use the Streamable HTTP transport, which means any client that supports this MCP transport can connect without additional configuration beyond the URL and authentication credentials.
Export Types
When you deploy a ToolSet, you choose an export type that determines how the tools are exposed:
| Type | Description |
|---|---|
mcp_server | Standard MCP Streamable HTTP endpoint. This is the primary deployment type and is recommended for most use cases. |
copilot_connector | Optimized for GitHub Copilot integration. |
sdk_bundle | Packaged for embedding in custom applications. |
http_catalog | REST-based catalog for tool discovery. |
For most scenarios, use the mcp_server export type. It provides the widest compatibility with MCP clients including Claude Desktop, Cursor, and other AI agents.
Environments
Each export is deployed to a specific environment, allowing you to isolate testing from production traffic:
| Environment | Purpose |
|---|---|
dev | Development and experimentation. Use this to test tools before promoting them. |
stage | Pre-production validation. Use this for integration testing with staging AI agents. |
prod | Live AI agent access. Production-grade reliability and monitoring. |
Each environment gets its own server key and endpoint URL. A single ToolSet can have exports in multiple environments simultaneously, so you can test changes in dev without affecting prod.
Authentication
MCP Servers support two authentication methods to control who can connect and invoke tools.
Service Key
Service key authentication is the default method. The MCP client includes a Nexla service key in the Authorization header of every request:
Authorization: Bearer <NEXLA_SERVICE_KEY>
This method is simple to configure and well suited for:
- Local AI agent setups (e.g., Claude Desktop on your machine)
- Server-to-server integrations
- Development and testing
See Service Keys for instructions on creating and managing service keys.
OAuth
OAuth authentication is designed for user-facing AI agents where individual user identity matters. Nexla supports OAuth with the following providers:
- Google -- Users authenticate with their Google account.
- Azure -- Users authenticate with their Microsoft Azure AD account.
With OAuth, the MCP client handles the full OAuth flow (redirect, user consent, token exchange). This is the recommended approach when:
- Multiple users share the same AI agent and you need per-user audit trails
- Your organization requires SSO-based access control
- You are deploying tools through a shared service like ChatGPT
Server Lifecycle
Every MCP Server export follows a lifecycle with four possible statuses:
| Status | Description |
|---|---|
draft | The export has been created but is not yet accepting connections. Use this state to configure settings before going live. |
active | The export is live and accepting connections from MCP clients. Tools are discoverable and invocable. |
paused | The export is temporarily disabled. Existing sessions may continue, but no new connections are accepted. Configuration is preserved. |
retired | The export is permanently deactivated. It cannot be reactivated. Use this when the server is no longer needed. |
The typical progression is draft to active for initial deployment, with paused available for maintenance or incident response, and retired for end-of-life.
Retiring a server is irreversible. If you need to temporarily disable an MCP server, use paused instead.
Next Steps
- Deploy an MCP Server -- Walk through deploying a ToolSet as an MCP server and managing the server instance.
- Connect MCP Clients -- Configure Claude Desktop, Cursor, and other AI agents to connect to your MCP server.
- MCP Protocol API -- Technical reference for the MCP protocol endpoints.