Updated
Agent building blocks

MCP Server for Your Product

Let AI agents — Claude, ChatGPT, Cursor and your users' own agents — use your product by exposing its actions and data as a Model Context Protocol server.

Ask your AI about this, with this page as the source:ChatGPT ↗Claude ↗Perplexity ↗

Trends

MCP Handler gained the most ground in package downloads: 8.4× against the median npm package over a year.

Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.

ToolDownloadsStarsProduct Hunt
MCP SDK52.9M/wk★ 13.6k—
Cloudflare21.3M/wk3.1× vs npm—7.8k followers
FastMCP12.2M/wk★ 28.1k—
MCP Handler1M/wk8.4× vs npm★ 686—
xmcp20.1k/wk★ 1.3k—

The real choice

Official SDK or a framework on top of it, and where the server runs. The official SDKs give you the protocol with nothing hidden, and you wire up transport, auth and deployment yourself. Frameworks add decorators or file-based tools, auth helpers and one-command deploys. A remote server runs next to your app — as a route in your web framework or on an edge platform — so users connect with a URL and sign in with OAuth instead of installing anything.

Pick by situation

Tap the ones that are you — the tools that fit light up below.
  • You want the protocol directly, in any major language, with no extra layerMCP SDK
  • Your backend is Python and you want tools defined with decorators, plus auth and deployment helpersFastMCP
  • You want the MCP server as a route in your existing Next.js or TypeScript appMCP Handlerxmcp
  • You want a remote MCP server on edge infrastructure, with OAuth and per-session state handledCloudflare

Compare the tools

Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.

ToolBest for
Official SDK1
MCP SDKlibrary · open sourceBuilding MCP servers and clients against the protocol directly, with official SDKs for TypeScript, Python, Go, Java, C#, Rust and more.508 open source52.9M/wk

Use it whenYou want full control over transport and auth, or work in a language frameworks don't cover.

Trade-offLow-level — auth, session handling and deployment are yours to build.

No maker's product we track shows it for this yet · in 508 open-source projects.

Framework3
FastMCPframework · open sourcePython MCP servers with tools defined by decorators, plus built-in auth providers, server composition and a testing client.72 open source12.2M/wk

Use it whenYour backend is Python and you want a working server in a few lines.

Trade-offPython only; its hosted option (Prefect Horizon) is aimed at teams and enterprises.

No maker's product we track shows it for this yet · in 72 open-source projects.

MCP Handlerlibrary · open sourceServing an MCP server from a route in a Next.js or other TypeScript app, on serverless functions.19 open source1M/wk8.4× vs npm

Use it whenYour product is already a Next.js app, especially one deployed on Vercel.

Trade-offA thin adapter over the official SDK, so tool design, auth and data access are still yours; the older SSE transport needs Redis.

No maker's product we track shows it for this yet · in 19 open-source projects.

xmcpframework · open sourceTypeScript MCP servers with file-based tool routing, hot reload and ready deploy targets.1 open source20.1k/wk

Use it whenYou want a standalone TypeScript MCP server with the conventions of a web framework.

Trade-offA young project with a smaller community than the official SDK.

No maker's product we track shows it for this yet · in 1 open-source project.

Edge platform1
Cloudflareplatform · free tierRemote MCP servers on Workers, with the Agents SDK handling per-session state and a provider library handling OAuth.No evidence yet21.3M/wk3.1× vs npm

Use it whenYou want a public, remote MCP server without running servers, or already host on Cloudflare.

Trade-offWorkers' runtime limits and Cloudflare-specific APIs, so the server is harder to move elsewhere.

llms.txtMCP

No maker's product we track shows it for this yet.

Before you choose

How to approach it

Start with three to five tools that map to the things users most often ask your product to do, not one tool per API endpoint. Write each tool's description for a model that has never seen your product, and return short, structured results instead of whole API payloads. Ship it as a remote server with OAuth so users connect with a URL; a local server that needs an API key pasted into a config file loses most people.

Common mistakes
  • Wrapping every REST endpoint as its own tool, so the agent drowns in near-identical options and picks the wrong one.
  • Accepting a long-lived API key in the server config instead of OAuth, so users paste secrets into files and can't limit what the agent may do.
  • Returning full database records from tools, filling the agent's context with fields it doesn't need and data users didn't mean to share.

Other options

Real choices most makers here won't need to weigh.