Skip to content

MCP basics

What is an
MCP server?

An MCP server is a small program that exposes a defined set of capabilities to any compatible MCP client, over a standard protocol. It is the thing being asked, not the thing asking.

Five steps, every call

The same handshake, each time.

Whether the underlying data is a calendar, a codebase or a memory store, the shape of an MCP call is the same. That consistency is the entire point of the protocol.

  1. 1. Connect

    A host opens a connection and asks the server what it can do.

  2. 2. Describe

    The server returns its list of capabilities — names, inputs and what each one returns.

  3. 3. Call

    The host, on the model's decision, calls one capability with specific arguments.

  4. 4. Check

    The server checks authorization and your privacy preferences before doing anything else.

  5. 5. Answer

    The server returns the smallest useful result — never a raw dump of underlying data.

Where a server runs

Servers can run on your machine, or somewhere else.

Some MCP servers run locally, as a process on your own computer that a host launches and talks to directly. Others run remotely, reachable over the network like any other hosted service. The Kelven MCP Gateway is a remote server: your memory stays where it already is, and a connected client reaches it over an authenticated connection rather than by downloading anything.

The tradeoffs between local and remote servers — and why remote is the right model for a memory layer — are covered in Remote MCP servers, explained.

In practice

What it will answer. What it won't.

  • Six capabilities

    Search memory, get relevant context, search projects, get project context, get user preferences, get privacy preferences — all read-only.

  • No write access

    A connected client cannot create, edit or delete anything in your memory through the Gateway.

  • Local-only respected

    Anything marked local-only never enters the Gateway's pipeline, regardless of which client is asking.

A server only ever does what it was built to expose — a badly scoped server can leak far more than intended, which is why the Gateway's capabilities are deliberately narrow and read-only rather than a general "give me everything" endpoint.

Related reading

Start from the concept, or the comparison.

If you haven't already, start with What is MCP? for the host/client/server picture, or see how this compares to a plain API in MCP vs. API. For how memory itself is protected, see the Privacy Center or the full MCP Gateway overview.

The future isn’t another AI model. It’s an intelligence layer that’s yours.

Kelven is the user-controlled intelligence layer between you and AI — with one memory that’s always yours.

No spam. Just one note when early access opens.