Remote MCP Server on Cloudflare

Remote MCP Server on Cloudflare avatar

by hufeng03

remote-mcp-servercloudflareBrowser Automation

What is Remote MCP Server on Cloudflare

Remote MCP Server on Cloudflare

Let's get a remote MCP server up-and-running on Cloudflare Workers complete with OAuth login!

Develop locally

# clone the repository
git clone [email protected]:cloudflare/ai.git

# install dependencies
cd ai
npm install

# run locally
npx nx dev remote-mcp-server

You should be able to open http://localhost:8787/ in your browser

Connect the MCP inspector to your server

To explore your new MCP api, you can use the MCP Inspector.

  • Start it with npx @modelcontextprotocol/inspector
  • Within the inspector, switch the Transport Type to SSE and enter http://localhost:8787/sse as the URL of the MCP server to connect to, and click "Connect"
  • You will navigate to a (mock) user/password login screen. Input any email and pass to login.
  • You should be redirected back to the MCP Inspector and you can now list and call any defined tools!

Connect Claude Desktop to your local MCP server

The MCP inspector is great, but we really want to connect this to Claude! Follow Anthropic's Quickstart and within Claude Desktop go to Settings > Developer > Edit Config to find your configuration file.

Open the file in your text editor and replace it with this configuration:

{
  "mcpServers": {
    "math": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "http://localhost:8787/sse"
      ]
    }
  }
}

This will run a local proxy and let Claude talk to your MCP server over HTTP

When you open Claude a browser window should open and allow you to login. You should see the tools available in the bottom right. Given the right prompt Claude should ask to call the tool.

Deploy to Cloudflare

  1. npx wrangler kv namespace create OAUTH_KV
  2. Follow the guidance to add the kv namespace ID to wrangler.jsonc
  3. npm run deploy

Call your newly deployed remote MCP server from a remote MCP client

Just like you did above in "Develop locally", run the MCP inspector:

npx @modelcontextprotocol/inspector@latest

Then enter the workers.dev URL (ex: worker-name.account-name.workers.dev/sse) of your Worker in the inspector as the URL of the MCP server to connect to, and click "Connect".

You've now connected to your MCP server from a remote MCP client.

Connect Claude Desktop to your remote MCP server

Update the Claude configuration file to point to your workers.dev URL (ex: worker-name.account-name.workers.dev/sse) and restart Claude

{
  "mcpServers": {
    "math": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://worker-name.account-name.workers.dev/sse"
      ]
    }
  }
}

Debugging

Should anything go wrong it can be helpful to restart Claude, or to try connecting directly to your MCP server on the command line with the following command.

npx mcp-remote http://localhost:8787/sse

In some rare cases it may help to clear the files added to ~/.mcp-auth

rm -rf ~/.mcp-auth

Leave a Comment

Comments section will be available soon. Stay tuned!

Frequently Asked Questions

What is MCP?

MCP (Model Context Protocol) is an open protocol that standardizes how applications provide context to LLMs. Think of MCP like a USB-C port for AI applications, providing a standardized way to connect AI models to different data sources and tools.

What are MCP Servers?

MCP Servers are lightweight programs that expose specific capabilities through the standardized Model Context Protocol. They act as bridges between LLMs like Claude and various data sources or services, allowing secure access to files, databases, APIs, and other resources.

How do MCP Servers work?

MCP Servers follow a client-server architecture where a host application (like Claude Desktop) connects to multiple servers. Each server provides specific functionality through standardized endpoints and protocols, enabling Claude to access data and perform actions through the standardized protocol.

Are MCP Servers secure?

Yes, MCP Servers are designed with security in mind. They run locally with explicit configuration and permissions, require user approval for actions, and include built-in security features to prevent unauthorized access and ensure data privacy.