Get in touch

MCP server development and training

We build Model Context Protocol (MCP) servers that let Claude, ChatGPT, Microsoft Copilot and other AI assistants work with your product, database or internal systems, under the same access rules your users already have. And we teach your engineers to build the next one.

Discuss your MCP project Read the case study

AI assistants

  • Claude
  • ChatGPT
  • Copilot

MCP server

  • search_customers
  • get_open_items
  • get_document
  • create_draft

OAuth 2.1, per-user permissions, audit log

Your systems

  • REST API
  • PostgreSQL
  • Search index
Protocol
MCP over Streamable HTTP (remote) or stdio (local)
Clients
Claude, ChatGPT, Copilot Studio, VS Code, Cursor, your own app
Auth
OAuth 2.1 with PKCE and dynamic client registration; per-user tokens; API keys between servers
Hosting
Your cloud or ours, or packaged for your own servers
Timeline
Typically 2–4 weeks for a read-only server over an existing API

When an MCP server makes sense

  • You run a data product or SaaS and your customers already work in Claude or ChatGPT. An MCP server puts your product inside their assistant instead of making them copy and paste.
  • Your staff need answers from live systems, such as the ERP, the CRM or the ticketing tool, not only from documents.
  • You want one integration for every assistant instead of a separate plugin for each vendor.

When it does not: if a nightly export would do, or nobody will reach the data through an assistant. We will say so.

How the pieces fit

The assistant never touches your database directly. It calls tools; each tool checks who is asking and returns only what that user may see.

  1. 1 AI clients Claude · ChatGPT · Copilot Studio · your web app
  2. 2 MCP server Tools with typed inputs · OAuth and per-user scopes · rate limits and audit log · tracing
  3. 3 Your systems REST or GraphQL API · SQL database · search index · ERP or CRM

The user’s identity travels with every call, so the server never has broader rights than the person asking.

What you get

  • Tool design

    A small set of well-named tools mapped to real user questions, with responses sized for a model’s context window.

  • Authentication

    OAuth 2.1 so each person connects with their own account and sees only their own data.

  • Evaluation

    A test set of real questions run against Claude and ChatGPT to check tool choice and answer quality before release.

  • Observability

    Every tool call traced, with latency, error and usage dashboards.

  • Deployment

    Containerised or serverless, in your cloud or ours, with CI and a staging environment.

  • Docs for your users

    Setup guides for each assistant your customers or staff use.

MCP training

Your team builds the next one.

A hands-on workshop for engineering teams: protocol basics, designing tools a model can use well (and the usual mistakes, shown on deliberately bad servers), authentication, testing with MCP Inspector, evaluation and deployment. Your engineers leave with a working server against one of your own APIs.

Questions

Which assistants support MCP?

Claude, ChatGPT, Microsoft Copilot Studio, GitHub Copilot in VS Code, Cursor and many others. The protocol is open and not tied to one vendor.

Is it safe to expose our data to an AI assistant?

The assistant only sees what a tool returns, and a tool only returns what the connected user may see. Rate limits and audit logs come as standard.

How many tools should a server have?

As few as cover the real questions. Our production servers range from a dozen to a few dozen tools; beyond that models start picking the wrong one, so we group tools, filter them per user or add a discovery tool.

Can we build it ourselves?

Often, yes. That is what the training is for. Many teams have us build the first server and review the next ones.