MCP Remote Connections
Remote MCP connections use the HTTP/SSE transport, allowing AI applications to connect to internet-hosted tools and data sources that reside outside the local environment.
Architecture
- Host: The AI application (e.g., Claude.ai) acts as the client.
- Server: A web-hosted service that implements the MCP protocol over HTTP.
- Bridge: Custom Connectors act as the secure link between the host and the remote server.
Connection Process
- URL Registration: Provide the host application with the HTTPS URL of the remote MCP server.
- Authentication: Remote servers typically require secure authentication, such as:
-
- OAuth: Redirecting the user to a third-party provider.
- API Keys: Provided during the setup phase.
-
- Capability Handshake: Once authenticated, the host and server negotiate features via the protocol.
Use Cases
- SaaS Integration: Connecting to project management tools (Linear), repositories (GitHub), or communication platforms (Slack) that are not local to the machine.
- Shared Infrastructure: Centralized servers that provide context or tools to an entire team or organization.
Security Considerations
- Encryption: All remote traffic must be encrypted via TLS/SSL.
- Authorization: Hosts can granularly enable or disable specific tools from the remote server.
- Trust: Only connect to remote servers from trusted sources, as they receive context from your conversations.
References
- Source:
00_Raw/mcp/Connect to remote MCP Servers.md - mcp-transport
- mcp-authorization
- mcp-local-connections
- mcp-architecture
- mcp-security
- mcp-client-development