How to Connect an LLM with MCP
MCP — the Model Context Protocol — has become the standard way to connect AI tools to outside data and services
MCP — the Model Context Protocol — has become the standard way to connect AI tools to outside data and services. If you've set up an MCP server before, you've probably connected an LLM to a database, a filesystem, or GitHub. Here's how the connection works, and a question most setup guides skip: once you've connected your LLM to MCP, what should you actually serve it?
How do you connect an LLM with MCP?
Through your tool's MCP configuration. The exact path depends on the client, but the shape is the same:
- Claude Desktop — add the server to
claude_desktop_config.json(amcpServersblock with the command or URL and any credentials), then restart. - Cursor — add the server via Settings → MCP, or a config file, in the same format.
- Claude Code — use
claude mcp addwith the transport and server details.
For remote servers you point at a URL and authenticate with a token; for local ones you point at a command that launches the server. Once connected, the LLM can see the server's resources and tools and use them during a session. That's the whole mechanic — connection is rarely the hard part.
What is MCP and what does it connect?
MCP is an open protocol — donated to the Linux Foundation in late 2025 — that standardizes how AI tools connect to outside systems. Before it, every tool needed a custom integration for every data source. MCP makes that a single common interface.
It connects two kinds of things: resources (structured data the model can read) and tools (actions the model can take). An MCP server exposes one or both. A GitHub server exposes your repos as resources and actions like opening a PR as tools. A filesystem server exposes files. The model, through an MCP-compatible client, reads and acts on whatever the server offers.
What can you connect an LLM to via MCP?
Almost anything with an MCP server: databases, filesystems, GitHub, Slack, Gmail, calendars, project tools. This is what most people set up — connecting the LLM to the services they use so it can fetch data and take actions.
But there's a category most people miss: your own structured context. MCP doesn't only connect tools and data sources. It can serve a structured profile of who you are — your role, background, current work, preferences — as a resource the LLM loads at the start of every session. Same protocol, different (and arguably more valuable) payload.
Connecting tools vs connecting context
Here's the distinction that matters, because it's easy to assume connecting your tools gives the LLM your context. It doesn't.
Connect Gmail via MCP and the LLM can search your email — but searching isn't understanding. Ask it to pull "everything about my Q3 strategy" and depending on how a connector retrieves, it may surface what literally matches and miss closely related items. A chat tool with connectors is a retrieval client, not an extraction engine. Connecting the source isn't the same as having your context.
Connecting context is different: a structured profile, already extracted and organized, that the LLM loads as a resource — not a raw source it has to search on demand. One tells the model where to look; the other tells it what's true.
So when you connect an LLM with MCP, connect your tools for actions and data — and connect a context layer for who you are. The first lets the model do things. The second lets it do them with your situation in mind.
→ How to load personal context via MCP: How to Load Your Personal Context into Claude and Cursor via MCP
→ How context delivery works: How to Deliver Personal Context to AI Tools
