Let the AI use your systems. Only the parts you choose, with a log of every call.
An MCP server is a small, standard connector that lets AI assistants such as Claude read from and act on your own software. We build custom ones for very specific needs: your point of sale, bookings, files and internal tools, with permissions and a record of everything the assistant does.
This is for you if
Three signs it is time.
- Your team already uses an AI assistant, and keeps pasting exports into it by hand.
- You want the assistant to look things up in your systems, but not to change anything without a person saying yes.
- No ready-made connector exists for the tools you run, or the ones that do expose far more than you want.
What changes
Answers from the real system
The assistant reads the live number from your system, not a spreadsheet someone pasted in last week.
Narrow by design
Each tool does one job, like looking up a store's sales or finding a booking. Nothing is exposed beyond what that job needs.
A record of every call
Every request is logged with who asked, what was read and what changed, so any answer can be checked later.
What's included
The usual scope. Every quote says exactly what is in and what is out.
Design
- The exact list of tools the assistant gets, in plain words
- What each tool may read, and which actions need a person to approve
- Test prompts that show the tools behaving, including ones that should be refused
Build
- A custom MCP server in TypeScript or Python
- Connections to your databases, APIs, files and internal apps
- Sign-in and permissions per person or team
- Read-only tools by default, with changes behind an approval step
Run
- Hosting on your servers or ours
- A log of every tool call
- Updates when your systems or the MCP standard change
How it runs
The general process, shaped for this kind of job.
- 01
List the jobs
We sit with the people who will use it and write down the questions and actions they want the assistant to handle, in their words.
- 02
Draw the boundary
For each job we agree what the assistant may read, what it may change, and what always waits for a person.
- 03
Build and test
We build the server, connect it to your systems and test it with real prompts, including the ones it should turn down.
- 04
Roll out and watch
We connect it to the assistants your team uses, read the logs for the first weeks and tighten anything that turns out too loose.
More of this kind of work. From the index.
Questions
Asked on first calls about this.
What is MCP, in plain terms?
Model Context Protocol is an open standard for connecting AI assistants to other software. A server offers a set of tools, such as "look up yesterday's sales for a store", and any assistant that speaks MCP can use them.
Which AI assistants can use it?
Any assistant that supports MCP, which includes Claude and a growing list of others. We check the ones your team actually uses before we start.
Can the AI change things in our systems?
Only where you decide it should. Tools are read-only by default. Anything that changes a record, spends money or messages a customer can be set to wait for a person to approve it.
Where does our data go?
The server runs on your infrastructure or ours, and only the result of each tool call goes to the assistant. Before launch we show you exactly what every tool returns.
Want this for your business?
Tell us what you have today and what is getting in the way. A person reads it and replies within two business days.