MCP server development for your product or your data
I build the MCP server that lets Claude, ChatGPT and other assistants use your product or query your data directly, with authentication, scoped tools and tests.
- You bring
- An API or a database, and what an assistant should be able to do with it
- You get back
- An MCP server, authenticated, tested, documented and listed
- How it is cut
- Two milestones: read tools live, then write tools and the listing
What you get
- An MCP server in TypeScript or Python, over streamable HTTP for remote use or stdio for local use.
- Tools designed for a model to call: few, clearly named, with descriptions and schemas that leave no room to guess.
- Authentication that fits your product, OAuth 2.0 with PKCE where users sign in, and a scope per tool.
- Tests that call every tool the way a client does, and a record of each call for audit.
- Documentation for your users, and the submission to the directories you want to be listed in.
How the work is done
Design the tools for the model
A REST API wrapped one endpoint per tool makes a poor MCP server. A model picks a tool from its name and description, so twelve vague tools perform worse than four exact ones. I start from the questions a user will ask the assistant and work back to the tools.
Read first, write second
The first milestone exposes read tools only, live, so you can use it from Claude and ChatGPT the same week. Write tools come second, each with a scope, a confirmation where the action cannot be undone, and a record.
Treat the assistant as an untrusted client
Whatever reaches a tool was written by a model that read text from somewhere. Every argument is validated, every tool has the narrowest permission that does the job, and nothing in a tool result is trusted as an instruction.
Where I have done this
The AIGROW MCP connector: Claude and ChatGPT query the AIGROW Score of a business, a 0 to 100 measure with a 95% interval, straight from a conversation.
aigrow.gg/aigrow-score(verifiable source, opens in a new tab)
Anima: a browser runtime where an assistant drives a signed-in browser through an MCP server, with real clicks and keystrokes.
The ReelVance engine exposed its whole pipeline to agents through an MCP server.
Who it fits
A good fit when
- You have a product or a dataset people would rather ask an assistant about than open a dashboard for.
- There is an API or a database to build on, even an internal one.
- You want it listed where people look for connectors, not only in a README.
Not a fit when
- There is no API and no stable data model underneath yet.
- The goal is a plugin for one assistant only and its vendor already offers a no-code route.
Questions about MCP servers
- How much does it cost to build an MCP server?
- The price follows the number of tools, whether any of them write, and how users authenticate. A read-only server over an existing API is a small project; a server with OAuth and write tools that can spend money or delete data is a larger one, because the gates and tests are most of the work. I quote per project after reading your brief.
- Do I need an MCP server if I already have an API?
- An assistant cannot use your API by itself: it needs tools it can discover, with descriptions it can choose between, over a protocol its client speaks. The MCP server is that layer. The API stays as it is underneath.
- Will it work with both Claude and ChatGPT?
- Yes. Both support remote MCP servers, and I test against both clients, because they differ in how they authenticate and in how they pick a tool from its description.
- Is it safe to let an assistant call tools on my product?
- It is as safe as the tools are narrow. Each tool gets the smallest permission that does its job, arguments are validated as hostile input, anything irreversible asks for confirmation, and every call is recorded. I do not expose a tool I could not explain to your security reviewer.