Asset 20 8 2
Does AI recommend your business? Run the free check →

Join 15,000 business owners, marketers and entrepreneurs. The Sunday newsletter you'll be annoyed only arrives once a week.

Article

MCP Server Builders: 9 Ways to Ship One in 2026

An MCP server builder is a framework, SDK or hosted platform that turns your code or API into a Model Context Protocol server that Claude, ChatGPT and other AI clients can call. Write it yourself with a Python or TypeScript framework, generate it from your API spec, or let a hosted platform deploy and secure it.

An MCP server is a small service that tells an AI client which tools it can use, what each one does and how to call it, all over the Model Context Protocol. If you own an API, a database or an internal system and you'd like Claude, ChatGPT or a coding assistant to use it directly, you need one, and the builders below are how you make it without writing the protocol plumbing by hand. Product teams use them to put their app inside AI chats, developers use them to give their own assistants access to private data, and API companies use them so customers can plug in without writing glue code.

Code your MCP server, generate it from an API, or hand it to a host

What you already have decides this more than any feature list. An empty repo, a documented API and a finished server each point somewhere different.

If the server is mostly new logic (search across your docs, a workflow that strings three internal calls together, a tool that writes to your CRM with checks first), write it with a framework or SDK. You write ordinary functions, add a description and typed inputs, and the framework handles the handshake, the tool list and the transport. This is where Python versus TypeScript matters. The protocol's own SDK page puts TypeScript, Python, C#, Go, Rust and Ruby in its top tier, so none of those leaves you on a side road. Pick the one your team already ships in; a server is code you'll maintain for years, not a weekend experiment.

Already sitting on a well-described REST API? Then writing every tool by hand is wasted effort. Generators read your OpenAPI spec and produce a server with one tool per operation, which you then trim. The trimming is the real work: in my view a spec with two hundred endpoints makes a poor MCP server, because the model is left choosing between two hundred near-identical tools. Expect to cut it to the dozen jobs your users ask for and rewrite the descriptions in plain words.

Hosting is the third question, and it arrives later than people expect. A local server that runs over standard input is fine for your own laptop. The moment customers or colleagues need it, you need a remote server over HTTP, with sign-in, logs and somewhere to run it. Hosted MCP platforms do that part, and some add analytics, cross-client testing and distribution. Serverless hosts you may already pay for also publish guides for running a remote server on their infrastructure.

Most teams end up mixing two: a framework to write the tools and a host to run them. That's fine. What you want to avoid is picking a host first and then discovering its SDK only speaks a language your team doesn't use.

Two shifts this year are worth knowing before you choose. Smithery shut its proxy and managed identity services on 31 October 2026 and now runs only as a public registry, so any guide that tells you to deploy there is stale. And Stainless, which sold MCP generation from OpenAPI specs, now carries a banner saying it has joined Anthropic. Frameworks are moving too: FastMCP's site announces version 4 for the new protocol, so check the date on any tutorial you follow.

Frameworks and SDKs for writing an MCP server yourself

FastMCP

If your team writes Python, this open-source framework handles the protocol for servers, clients and interactive apps so you only write the tool logic.

xmcp

A TypeScript option that turns plain functions into MCP tools, with file-system routing, starter templates and deployment from the Vercel command line.

MCP Framework

Open-source TypeScript framework with a project generator, automatic tool discovery and built-in support for OAuth and HTTP streaming.

MCP TypeScript SDK

The official, open-source TypeScript SDK, for developers who'd rather write servers and clients straight against the protocol with no extra layer.

Platforms that generate, deploy or host your MCP server

Manufact (mcp-use)

Teams taking an MCP app to production get an SDK plus hosting, cross-client testing, analytics and a cloud inspector in one platform.

Alpic

A cloud host built for MCP servers and ChatGPT apps, deploying in one click and adding analytics, distribution and open-source tooling.

Speakeasy

API companies can generate a server from their existing API, then host and secure it with enterprise sign-in and per-customer logs.

Cloudflare Agents

Serverless route for running a remote MCP server on Cloudflare's network, deployed from the dashboard or CLI with an optional OAuth sign-in.

Work with me

Want AI doing the heavy lifting in your marketing?

I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.

Vercel

Already deploying web projects here? Its mcp-handler package lets you build an MCP server and ship it like any other project.

Sign-in, permissions and what your MCP server lets a model touch

The protocol's own specification sets the ground rules here, and it's worth reading its authorization and security best-practice pages before you choose a builder, because a builder that ignores them leaves you to do the work.

Sign-in is optional in the specification, but how you do it depends on the transport. A server that talks over HTTP should follow the spec's authorization rules, which are built on OAuth 2.1, with your server acting as the resource server that checks access tokens. A local server running over standard input shouldn't use that flow and should take its credentials from the environment instead. The security guidance is blunt about one shortcut: a server must not accept tokens that weren't issued for it, so passing a user's existing API token straight through is out.

Local servers get their own warning. The best-practice page points out that a server on someone's machine may have direct access to their system and may be reachable by other processes, so a local server is less harmless than it feels.

On permissions, the same page recommends a least-privilege scope model: start with a small set covering low-risk discovery and read operations, and ask for more only when a privileged action is first attempted. The tools section adds the server's duties: validate every input, enforce access controls, rate limit calls and sanitise outputs. On the client side it says there should always be a human able to deny a tool call, with confirmation prompts for sensitive operations, tool results checked before they reach the model, and tool usage logged for audit.

My own reading of all that: ship read-only tools first, add each write action on its own with a name that says what it changes, and keep a log of who called what with which inputs, because customers who connect your server will ask. If you're putting several servers and models behind one control point, an AI gateway can centralise keys, rate limits and logs instead of rebuilding them in every server.

Do you make MCP server frameworks, SDKs or hosting platforms? Tool makers can ask to be considered for a featured place on this list.

What an MCP server costs to build and run

Writing the server is mostly your own time, since the frameworks and the official SDK here are open source. Bills start once the server runs, and hosts meter in different ways: some count requests, with a free monthly allowance and paid tiers that raise it; others sell usage credits, with a few dollars included free; the gateway style bills by the volume of data flowing out. Cloudflare and Vercel don't list an MCP-specific price, so a server there falls under whatever plan you already run. The bigger cost is upkeep: every protocol revision and every change to your API means updating tools and descriptions.

Where to start with your first MCP server

If you're a Python shop building something new, start with a Python framework and run it locally until the tools feel right in a real client, then move it to a host. TypeScript teams have the same choice between a framework with routing and generators or the bare official SDK; take the framework unless you need control over every message.

If you sell an API and want customers using it from AI clients, a generator with hosting and per-customer logs saves the most work, as long as you spend the time cutting the tool list down. If you want analytics, testing across clients and help with distribution, go for a hosted MCP platform. And if the real need is to connect your assistant to apps that already exist rather than build a server of your own, an AI agent integration platform will get you there faster.

Before you commit to any of these, give it an hour. Wrap one read-only call you know well (an order lookup or a docs search), run it locally, connect it from a desktop AI client and check the tool list shows exactly one tool with a description a stranger would understand. Then ask one question that should trigger it and one that shouldn't, and read what came back. If the builder made that hour painful, it won't get easier at thirty tools.

The mistakes I'd steer you away from are dull ones: following a tutorial written for an older protocol version, deploying somewhere that has since closed its hosting, exposing every endpoint because the generator made it easy, and adding write actions before the read tools behave. I'd ship three good tools before thirty average ones, and learn what people ask for before building the rest.

Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.

Published and maintained by the Lilach Bullock team, covering marketing, AI and business growth.
Your buyers are asking AI who to use. Does it say you?

See for free whether ChatGPT, Claude, Perplexity, Gemini and Google name you, and get the plan to become the answer.

Check my AI visibility →
Sundays only

Get the Sunday newsletter.

One email a week. AI experiments, marketing tactics, and the workflows Lilach is building right now in her own business.

Subscribe free

Let’s get your marketing running on AI.

Book a free 30-minute call

We figure out what you need, where AI fits in, and what working together would look like.

Book the call →

Or take the 30-second calculator

You’ll see the hours and the money quietly leaking out of your week, and the three workflows worth building first.

Take the calculator →

Or grab the free AI resource library

Prompt packs, templates, checklists, and swipe files. The exact tools I build for paying clients. Yours, free.

Get the library →
Keep reading

More from the blog.