Skip to content

Connector builds / Hire an MCP developer

MCP server development, built to pass review.

If you are hiring an MCP developer, hire for review, not for code. Any good engineer can wrap an API in an MCP server. What gets a server listed is everything around it: OAuth 2.0 so each customer signs in to their own account, read tools split from write tools, a title and a readOnlyHint or destructiveHint on every tool, errors a model can act on, a fully populated test account, and docs and privacy pages a reviewer can read. I build the whole surface on your public API, and the code lands in your repository.

Nine of our tools are listed in Anthropic’s Claude directory.

At a glance

Transport
Remote, over HTTPS, the only kind Anthropic’s directory accepts for a connector
Auth
OAuth 2.0 for anything that touches a customer’s account
Tool shape
Separate read and write tools, each with a title and a hint
You keep
The code, the hosting account and the eval set

Checked October 12, 2026

What an MCP developer should hand you

A server, and the paperwork reviewers read.

An MCP server is a small program that tells Claude or ChatGPT which actions your product can take, and then takes them through your API. The protocol is open and the SDKs are good. The hard part is judgment: which of your endpoints become tools, how each one is named and described, and which ones must stop and ask the customer before they write anything.

Here is what I deliver on a Connector Build, every time:

  • Six to twelve tools mapped to the jobs your customers come to you for, each one tied to a real endpoint in your docs
  • OAuth sign-in, so every customer connects their own account and sees only their own data
  • An approval step on every write: the tool drafts the change, the customer says yes in the chat, then it runs
  • Errors that say what went wrong and what to try, never a bare “Bad Request”
  • An eval set your team can rerun before every release
  • A docs page and a privacy page, written for the reviewer who reads them first
  • The submissions to Anthropic’s Claude directory and to ChatGPT

What gets a server rejected

The four rejections I design out first.

Anthropic publishes its review criteria, and they read like a list of shortcuts developers take. These four are the ones I rule out from the first sketch.

  • One catch-all api_request tool with a method parameter. Anthropic rejects any single tool that accepts both safe and unsafe HTTP methods; reads and writes have to be separate tools.
  • Missing annotations. Every tool needs a title plus readOnlyHint or destructiveHint, and those hints decide what Claude runs without asking and what always prompts the customer.
  • Descriptions that steer the model. A tool description that tells Claude how to behave, or to call tools the user did not ask for, is treated as prompt injection.
  • Generic errors and oversized responses. Every tool must succeed on valid input, and a summary request should not return a database dump.

How to judge an MCP developer

Ask to see something already listed.

MCP is young, so most portfolios are demos. Ask for three things: a server that is listed in a directory today, a connector that uses OAuth against real customer accounts, and an example of a write that waits for approval. If a developer cannot show all three, you are paying for their first one.

Ours are public. You can add any of our listed connectors to Claude by URL from Built for Claude, read the tool list each one exposes, and watch Owner’s Desk hold a drafted invoice until the owner confirms it.

In review with Anthropic
4
In the official MCP Registry
15
Tools Claude can call
69

What we have shipped

Built and running, not a deck.

See every connector we have built for Claude

The package that fits

Connector Build

One product, one connector, both directory submissions. If your own engineers will write the code, Connector Blueprint hands them the design instead.

Your customers do their real work in your product from inside the assistant they already use. Set scope, set price, quoted in the conversation, and changes to what we built are included.

Straight answers

What teams ask next.

What does an MCP developer actually build?

A remote server that lists your product’s actions as tools and runs them through your API when Claude or ChatGPT calls them. Around it: OAuth sign-in, an approval step on writes, an eval set, docs, a privacy page and the directory submissions.

Do you need access to our codebase?

No. I design from your public API docs and build against your API with the scopes you grant. If a job your customers want needs an endpoint you do not have yet, the proposal says so plainly.

Which language do you build in?

TypeScript on the official MCP SDK, unless your team runs something else and wants to own it in that language. The code goes in your repository either way.

Will the same server work in Claude and ChatGPT?

Yes. Both speak MCP, so one server can serve both. Each directory has its own listing, its own review and its own extra requirements, and I prepare both.

Sources

I wrote this page for Modern Mustard Seed, and we sell the service it describes. Every fact about Anthropic, OpenAI or another company was read from the page linked here on October 12, 2026.

One next step

Send me your API docs.

Tell me your product and where your API reference lives. I will come back with the jobs your customers would do in Claude and ChatGPT, the tools I would build, and the package that fits. Managed for you, end to end, and the code is yours.

Claude is a product of Anthropic, and ChatGPT is a product of OpenAI. Modern Mustard Seed is an independent studio and is not affiliated with or endorsed by either.