Custom MCP servers that connect Claude and ChatGPT to your systems
Knowlytica designs and builds custom MCP servers for organisations across Europe, Switzerland included. An MCP (Model Context Protocol) server gives Claude, ChatGPT, Cursor and your own agents a defined set of tools on top of your CRM, ERP, document stores and internal APIs. Every tool is read-only by default, write actions wait for a person’s approval, and every call is logged. The server runs in your cloud, on your servers or in a Swiss data centre; work is quoted per scope after a free 30-minute call.
Status: available now. Free 30-minute call, then a written proposal for the agreed scope.
How to connect your internal tools to Claude with an MCP server
Six steps, in the order we run them. Menu names are Claude’s as of October 2026; ChatGPT and Cursor use the same server through their own settings.
- Step 01
List the jobs, not the endpoints
Pick three to five tasks people repeat every week, such as finding the latest contract for a client. Each task becomes one tool with a verb-first name like
search_contractsorget_invoice_status. You get a one-page tool list with inputs, outputs and the system each tool reads. - Step 02
Mark every tool read or write
Tools that only look things up are marked read. Tools that create, change or send are marked write: they return a draft and wait for a named person to press Approve before anything changes.
- Step 03
Choose how people sign in
Use OAuth so each person reaches only what their own account allows. A fixed API key in Request headers is shared by every member of a Claude Team or Enterprise workspace, so keep it for narrow, read-only data.
- Step 04
Host it where Claude can reach it
We deploy the server over HTTPS in your cloud, on your servers or in a Swiss data centre. Claude’s hosted apps connect from Anthropic’s cloud, so a server behind a VPN or firewall needs Anthropic’s published IP ranges allowlisted.
- Step 05
Add it to Claude
A workspace Owner opens Organization settings › Connectors › Add, chooses Custom › Web, enters a name and the server URL, picks the Authentication option and clicks Add. Members then open Customize › Connectors and click Connect. In a chat it is switched on under + › Connectors; when it works, Claude shows which tool it called.
- Step 06
Test with real questions, then roll out
Before release we run 20 to 30 questions your team actually asks and check every tool call in the audit log. A wrong tool choice is fixed in the tool’s description, not by adding another tool.
Tool descriptions decide what the model calls
The model picks a tool from its name and its one-line description, never from your API documentation. A server that mirrors twenty endpoints with names like get_v2_records makes Claude guess: it calls the wrong tool, or chains five calls where one would do.
We write every description as one sentence a new colleague would understand, say when not to use the tool, and keep each server to the jobs of one team. When the list grows past that, it becomes a second server.
What Knowlytica adds to the MCP standard
MCP defines how a client and a server talk. It leaves out what an organisation needs to run one safely; that is what we add.
Read-only by default
Write tools return a draft and wait for a named person’s approval.
An audit log of every call
User, tool, inputs, result and time, exportable for your data protection officer.
One server, every client
The same server works in Claude, Cursor, your own agents, and ChatGPT where custom MCP apps are enabled.
20–30 test questions per release
Real questions from your team, rerun before every change ships.
Swiss or your own hosting
Your cloud, your servers or a Swiss data centre. Client data never trains public models.
A full handover
Code, tool descriptions, runbook and training, so your team runs it without us.
Several of our builds run on MCP: Flow, where agency teams drive a production canvas from Claude Desktop; Pigeonhole, which coordinates AI agents across a team; AgencyOS, an agency’s knowledge base that staff also query from Claude; and Global Vision, whose travel risk reports draw on more than 40 OSINT tools. All are on the work page.
Frequently asked questions
What is an MCP server?
MCP (Model Context Protocol) is an open standard Anthropic introduced in November 2024 for connecting AI apps to tools and data. A server publishes a list of tools; an AI client such as Claude, Cursor or ChatGPT reads the list and calls the tools it needs.
How much does it cost?
The first 30-minute call is free. Each server is quoted in a written proposal for the agreed scope. It runs on your own infrastructure, so hosting is billed by your cloud provider, and there is no licence fee from us.
What isn’t an MCP server?
It is not a chatbot and not a copy of your data. It stores nothing and trains no model; it passes each request to systems you already run, under the permissions they already have.
Does it work with ChatGPT and Cursor as well as Claude?
Yes. Claude and Cursor support MCP, and ChatGPT connects custom MCP servers on the plans where OpenAI enables them. One server serves all of them.
Claude says it can’t reach our connector. What happened?
Claude’s hosted apps connect from Anthropic’s cloud, not from the user’s laptop, so a server behind a VPN or firewall is unreachable until Anthropic’s IP ranges are allowlisted. If the server URL changed, remove the connector and add it again: its settings cannot be edited in place.
Can the server stay inside our network?
Yes, as a local server: Claude Desktop and Claude Code can run an MCP server on the user’s own machine, which reaches internal systems without exposing them to the internet. Claude on the web and on mobile needs a server it can reach from Anthropic’s cloud.
Also works with
This page covers MCP servers. To roll Claude out to a whole team, see Claude for business. For the legal side, see revFADP compliance for AI assistants and the EU AI Act for Swiss companies. Agencies can see MCP builds in use on AI agents for marketing agencies, and every service is on the Knowlytica home page.