MCP basics
What is MCP,
in plain language?
Model Context Protocol (MCP) is an open standard, introduced by Anthropic in late 2024, that lets an AI assistant call out to external tools and data during a conversation, using one consistent interface instead of a different integration for every app.
The problem it solves
Before MCP, every integration was custom.
If an assistant wanted to read your calendar, search a codebase, or look up a customer record, someone had to write a bespoke connector for that exact assistant and that exact data source. Multiply that by every assistant and every data source, and most useful integrations never got built at all.
MCP standardises the connection instead of the content: any MCP-speaking host can talk to any MCP-speaking server, the same way any browser can open any website.
One protocol, three roles
Host
The AI application you're actually talking to — ChatGPT, or another MCP client. It decides when to reach for a tool.
Client
The piece inside the host that speaks MCP and manages a single connection to one server.
Server
The program that exposes capabilities over MCP. The Kelven MCP Gateway is a server: it answers narrow, read-only questions about your memory.
Independence
One memory, any compatible app.
Kelven app
Complete experience
ChatGPT
Supported MCP client
Other MCP clients
Compatible
The intelligence layer
Kelven
Routing, memory and privacy rules live here — not inside any one interface.
Kelven controls what Kelven shares. It can’t hide something you typed directly into another app.
Capabilities, not data dumps
Servers expose actions, not databases.
An MCP server publishes a small menu of named capabilities — things like "search memory" or "get relevant context" — along with a description of what each one needs and returns. A host reads that menu and decides when to use it. The server still decides, on every single call, what it's willing to answer.
Common misreadings
Three things worth being precise about.
Not a new AI model
MCP carries requests and responses between a host and a server. The intelligence still comes from whichever model answers.
Not a universal login
Connecting one MCP client doesn't connect every AI you use. Each client has to add support, and support has to be verified for each one.
Not full access by default
A server decides exactly what it exposes. Kelven's Gateway exposes six read-only capabilities — nothing about your account is handed over wholesale.
Not a prompt filter
MCP cannot rewrite a message before the AI app receives it. Kelven only controls the memory context its own server returns.
Where Kelven fits
One memory, reachable from compatible clients.
Kelven's own MCP server is called the MCP Gateway. It lets a compatible MCP client ask narrow questions about your memory — never the memory itself — after checking what you've allowed. Kelven works with compatible MCP clients; it does not work with every AI, and it never will claim to.
Read the MCP Gateway overview for the six capabilities and the data-minimisation pipeline, the MCP server explainer for how a server like the Gateway is built, or the Privacy Center for how memory, retention and provider processing work in general.
Your AI, your context. Start with Kelven today.
Put powerful AI to work with memory you control and protection before cloud requests.