Custom MCPs
Custom MCPs for the systems you already run.
A Model Context Protocol server that gives an AI tool named actions against your data. Clear limits on what it can read, and what it can change.
A connector with a boundary.
MCP is an open way for tools such as Claude to call software you control. The server advertises tools. Each tool has a name, a description and a shape for its inputs. The assistant can call those tools. It does not receive a standing login to the rest of the business.
Keen Concepts builds custom MCPs when the system that matters has no suitable connector, or when a public connector is wider than you will accept. The protocol is the interface. The consultancy is deciding which actions belong on it. A plain-language account is in what an MCP server is.
What a build includes.
- The systems and records the job actually needs. Often a CRM, a document store, or an internal database.
- Tools written as specific actions: find a job, read a file, draft a note, file a result.
- A split between read-only tools and tools that change data.
- A place for a person to approve consequential actions.
- Documentation of what was built, and a pilot with the people who will use it.
If two systems need to exchange records on a schedule, with no assistant in the middle, that is an integration. We still do that work. MCP is the right shape when a person, or a tightly scoped agent, needs to ask and act across systems during the job.
Questions we expect.
What is a custom MCP?
A server built for your systems. It offers a short list of actions, such as finding a record or adding a note, instead of handing an assistant a general login.
When is a ready-made connector enough?
When it already reaches the system, does the actions you need, and your review accepts how it signs in and what it can change. A custom server is for the gap, or for tighter limits.
Does an MCP give an assistant access to everything?
It should not. The server exposes named tools, separates reading from changes, and leaves consequential actions for a person.
Where the model itself is allowed to run is a separate decision. See private and local AI when the data cannot leave a controlled environment.
Name the system.
Tell us which system the assistant cannot reach today, and what it should be allowed to do there.