The Model Context Protocol is how AI coding tools grow new hands. Claude Code, Cursor, Claude and others connect to MCP servers, small programs that expose tools, and through them the assistant can do things it could not on its own: query a database, call an internal API, read the filesystem, or reach out to the web. That last category is the one this page is about, because an MCP server that fetches the web is a web client, and web clients that run at scale need proxies.
Not every MCP server does. The proxy question splits cleanly, and knowing which side a server is on saves you from either a pointless purchase or a wall of blocks.
Which MCP servers need a proxy
The line is what the server reaches out to.
- Fetch-heavy servers need proxies. A web-fetch server, a scraping server, a search server, a price checker, a server that reads public pages or pulls listings, makes ordinary HTTP requests to sites that meter by IP, fingerprint the connection and block the address that asks too much. These need residential proxies in exactly the places any scraper does.
- Private-target servers do not. A server that talks only to your own database, an internal API, a paid data API with a key, or the local filesystem, never touches the open web as an anonymous client, so a proxy adds nothing.
There is a detail that pushes fetch servers toward proxies harder than a laptop scraper: an MCP server usually runs on infrastructure, a VPS, a container, a cloud function, so its default address is a datacenter IP, which is the exact profile many sites distrust on sight. A fetch server on a cloud box often gets challenged where the same request from a home connection would not. A residential exit is what makes its requests read as ordinary, which our explainer on how websites detect proxies explains.
Where the proxy goes
In the server, not in the client. This is the point people miss. The AI tool, Claude Code or Cursor, calls the MCP server and receives a result; it never sees how the server got that result. So the proxy lives in the server's own HTTP client, configured the same way as any scraper in that language. The client stays oblivious, which is correct: the proxy is the server's business.
Claude Code / Cursor
calls the MCP tool; never sees the proxy
MCP server
your program, exposes fetch tools
Proxy in the server's HTTP client
rotating residential or sticky ISP
The target site
sees an ordinary home address
Match the type to what the server fetches: rotating residential when it hammers a defended site, a sticky ISP exit when it keeps a session, datacenter for friendly open APIs. The exact code is the same as any scraper in that runtime, and our library guides cover the common ones: Node.js and node-fetch for a TypeScript MCP server, Python requests and httpx for a Python one, and Playwright or Puppeteer if the server drives a headless browser.
Building one with an AI coding tool
The neat part is that the tool calling the MCP server can also build it. Ask Claude Code or Cursor to write the MCP server and to route its fetch tools through a proxy from the start, giving it the gateway endpoint and the library pattern for your language, and it produces a server with the proxy already in the HTTP client. That is the fastest honest path to a fetch server that survives contact with a defended site, and it is the same integration our proxies for Claude Code page describes for any scraper the tool writes.
One boundary, so nobody wires it wrong: the proxy on the server changes only the requests the server makes to websites. It does not touch the AI tool's model quota or billing, which follow the tool's own account and key. The proxy is about the web the server reaches, full stop.
Where HProxy fits
For fetch-heavy MCP servers, point the server at the residential pool from $0.44/GB on defended targets, or a static ISP address at $2.70/IP a month when it keeps a session, metered by traffic with no KYC and sized to what the server pulls. Because MCP servers run on infrastructure, the move from a default datacenter IP to a residential exit is often the single change that turns a challenged server into a working one. Confirm the exit with the proxy checker, keep the proxy in the server's HTTP client, and the tools you expose to Claude Code or Cursor start returning data instead of block pages. The wider picture is in proxies and AI coding tools.
Sources
- Model Context Protocol, the open standard for connecting AI tools to servers that expose tools (modelcontextprotocol.io).
- HProxy, proxies for Claude Code and proxies for web scraping, plus the runtime library guides linked above.
- HProxy, how websites detect proxies: why a server's default datacenter IP gets challenged.