Building Remote MCP Servers
To get started with remote servers, start with the following resources:
The auth spec, especially details on the auth flow for third-party services
The remote server examples in the TypeScript and Python SDKs
The client and server auth implementations in the TypeScript and Python SDKs
The official MCP roadmap and draft spec’s changelog for details on how the protocol will evolve
Other resources (like this) may also be helpful to learn about considerations when building, deploying, and troubleshooting remote servers.
In addition, some solutions like Cloudflare provide remote MCP server hosting with built-in autoscaling, OAuth token management, and deployment.
Claude.ai MCP Support
Platforms
Remote MCP servers are supported on Claude.ai & Claude for Desktop for the Pro, Max, Teams, and Enterprise tiers.
To configure Remote MCP servers for use in Claude for Desktop, add them via Settings > Integrations. Claude for Desktop will not connect to remote servers that are configured directly via claude_desktop_config.json.
Claude’s mobile apps will add support for Integrations in the coming weeks
Transport & Auth
Claude supports SSE-based servers, with Streamable HTTP support coming soon.
Claude supports both authless and OAuth-based remote servers
For OAuth-based servers, Claude requires Dynamic Client Registration support and does not yet support a way for users to specify a client ID or secret.
Claude supports the 3/26 auth spec
Note: There are updates to MCP auth proposed in the latest draft specification. Claude will be adding support for this after it lands into an official spec revision, but doesn’t have an official timeline yet for doing so.
Protocol Features
Claude supports tools, prompts, and resources.
Claude supports text-, binary-, and image-based tool results
Claude supports text- and binary- based resources
Claude does not yet support resource subscriptions, sampling, and other more advanced or draft capabilities
Testing Remote MCP Servers
To test and validate your server, use the inspector tool. This will allow you to validate:
that Claude.ai and other MCP clients can successfully connect to your server
that your server successfully initiates and completes the auth flow
that your server correctly implements various parts of the auth flow
which tools, prompts, resources, and other MCP features your server exposes
See the MCP documentation for more details on using inspector and for other tips on how to debug and troubleshoot your server.
In addition, other solutions like Cloudflare’s AI Playground allow you to test remote MCP server functionality.