AI on your custom software
You have custom-built software—your own or a vendor’s—that runs the business, and you want AI to work inside it, not in a separate window. We connect it to your data and your processes, with controlled access and without rewriting what already works.
MCP · Claude · Amazon Bedrock · Node · .NET · AWS
What we do
A connector on top of your data
A custom MCP server that exposes what your software knows—customers, products, case files—to the AI, only for the people who should have access.
Natural language queries
The team asks and gets the answer without clicking through the interface. The question reaches the data by the short route.
Agents that take action
Not just looking things up: creating, updating, sending. Whole tasks delegated to an agent, with a log of who did what.
Review of the existing code
Before connecting anything, we look at where AI fits and where it doesn’t, and what needs reinforcing so it holds up.
What it looks like in practice
- Staff at a healthcare company ask about a patient and the answer comes from their backend, without opening the interface.
- A sales team searches a catalog of thousands of items from the conversation: search, compare, see the details.
- A process that used to mean opening three screens is done by an agent, and it gets logged.
What you need
- Access to the code or, at the very least, to the software’s API and database.
- A clear picture of who can see what: the AI inherits permissions, it doesn’t bypass them.
- A concrete use case, with someone who uses it every day and can say whether it works.
A case
Digital health company
- What was there
- Their data in their backend, and clinical staff who had to open the interface and click through it to look anything up.
- What we did
- A connector on top of their backend with credential-based access: staff ask and get the answer without opening anything.
How do you connect AI to custom-built software?
Sanfeliu builds a connector—a custom MCP server—on top of a company’s own software, so the AI can query and act on its data from the inside, with controlled access. The team asks in plain language and the answer comes from the real backend; an agent can carry out whole tasks (create, update, send) with a log of who did what. Before connecting anything, we review the existing code to see where the AI fits and what needs strengthening.
- Technology: Model Context Protocol (MCP), Claude, Amazon Bedrock; backends in Node, .NET, or whatever you already run.
- Requirement: access to the code or, at a minimum, to the software’s API and database.
- The AI inherits the software’s permissions: it doesn’t bypass them.
- First step: a free fit session; then an audit with a fixed quote.
- Cases: a connector on the backend of a digital health company, and another on the live catalog of an industrial distributor.
Questions we get asked
- What is an MCP server?
- A piece of software that tells the AI what it can look up and what it can do with your data, and that carries out those actions with the permissions of whoever is asking. It is the open standard that Claude, ChatGPT, and other assistants use to connect to business systems.
- Our software was built by another vendor. Can it still be done?
- Yes, if you have access to the code, or to the API and the database. If the vendor needs to be involved, we coordinate with them. The audit sets out what is possible and what isn’t.
- How much does it cost?
- It depends on what the AI needs to be able to do, not on hours. After the free fit session, the audit gives you a fixed quote for each feature.
- Is it safe to give AI access to our data?
- The AI has no direct access to anything: everything goes through the connector, which applies the user’s permissions, limits what can be done, and logs every action. The data goes to providers under enterprise agreements that don’t train on it.
- Does the software need to be rewritten?
- No. The connector is added alongside what is already there. If, while reviewing the code, we find things that need strengthening for it to hold up, we say so up front and quote that separately.
How we work
Let’s start by seeing if we’re a fit.
Eight questions and a call. Free, and we tell you what we think even if the answer is no.
It’s a call with the engineer who would lead the project, not with a salesperson.