MCP under the hood: hosts, clients & discovery
The optional-but-satisfying technical layer. What host, client and server really mean, why 'local' doesn't mean 'offline,' and how to find a trustworthy server in a noisy marketplace.
This page is the optional-but-satisfying technical layer. You don't need any of it to use MCP — but these are terms you'll hear constantly, and a couple of them cause real confusion that's worth clearing up now.
01Host, client, server
Three words, one relationship. Get this and the jargon stops being scary.
| Term | What it is | Example |
|---|---|---|
| Host | The overall AI app that needs to call tools. | Claude Code, ChatGPT, an OpenAI Agents SDK app. |
| Client | Code inside the host — one per set of tools you use. Baked in; you don't manage it. | The connector Claude Code spins up for each server. |
| Server | The thing you actually care about — manages the tools someone else wrote. | A web-browsing server; a stock-price server. |
When someone says "I'm using this MCP," they almost always mean "this MCP server." You hand your host a description of a server; it knows how to set up a client, launch the server and run it.
02Local vs remote — and two confusions to kill
Servers run in one of two modes (the "transports"). Neither is more "real" than the other.
Runs on your computer
Installed via uv/pip (Python) or npx (Node) from a shared repo. Most servers are local.
Runs in the cloud
Hosted by the provider (Atlassian for Jira, GitHub, etc.). Two techniques: streamable HTTP (current) and SSE (deprecated but still works).
A local server runs on your machine, but its tools may still make web calls. The stock-price server runs locally yet clearly fetches prices from the internet. Where the server runs ≠ whether it reaches the network.
People assume a local server isn't "proper" sharing. Wrong. The code was still fetched from a shared, remote repository and installed on your box. You get every benefit of MCP; it just executes locally.
03Discovery — the wild west
Here's the odd part: MCP gave us a great way to connect tools but never converged on one place to find them. There are thousands of servers across many sites, and navigating them takes real research.
| Where to look | What it is | Feel |
|---|---|---|
| registry.modelcontextprotocol.io | The official registry (MCP is now governed by the Linux Foundation). | Authoritative but quiet — "some crickets." |
| Anthropic's reference servers (GitHub) | Early first-party servers + "blessed" third-party ones. | A trusted place to start. |
| mcp.so | Marketplace with search, tags — ~17,000 servers. | Usually OK; noisy. |
| Glama.ai | Marketplace with security/quality ratings, downloads, favourites. | More thriving; nicer signals. |
Searching "Context7" on any marketplace returns a pile of look-alikes — unofficial clones, dubious forks. The reliable move is the same as any open-source software: go to the source on GitHub. Check the author (Context7 is by Upstash), the star count (40,000+), recent commits ("updated 10 hours ago"), and the README's own install instructions. That's your authoritative check.
✓ Key takeaways
- Host = the app; client = per-server code inside it; server = the tools someone wrote (the part you care about).
- "Using an MCP" almost always means "using an MCP server."
- Servers run local (most) or remote (streamable HTTP now; SSE deprecated).
- Local ≠ offline, and local is still fully "real" MCP — the code came from a shared repo.
- Discovery is scattered — mcp.so, Glama, the official registry — so verify on GitHub: author, stars, recent activity.