# What is an MCP server? Definition, architecture and real examples An MCP server is a program that exposes one product's data and actions to AI assistants through the [Model Context Protocol](https://mcpyet.com/what-is-mcp.md) — an open standard, so a single server built for Notion or GitHub works in Claude, Cursor, ChatGPT and every other compatible client. Think of it as an adapter. The product already has an API meant for developers; the MCP server re-presents that API in a form an AI model can discover and call on its own. The assistant asks the server what it can do, gets back a typed menu, and picks from it mid-conversation — no bespoke integration per app. ## What a server actually provides A server exposes up to three kinds of capability, and clients discover all of them by asking: - **Tools** — named actions with typed inputs: `create_issue`, `run_query`, `send_message`. The model decides when to call one; the user approves the calls that matter. Nearly every server ships tools; many ship only tools. - **Resources** — readable data the client can pull into context: a file, a database schema, a document. For giving the model something to read rather than something to do. - **Prompts** — reusable prompt templates the server offers to the user ("summarize this incident"), pre-wired to the server's own data and tools. ## The architecture MCP has three roles. A **host** is the AI app you actually use (Claude, Cursor, VS Code). Inside it, the host runs one **client** per connection, and each client talks to exactly one **server**: ``` ┌─────────────── host (Claude, Cursor, VS Code…) ───────────────┐ │ │ │ MCP client ──── MCP client ──── MCP client │ └───────│──────────────│───────────────│───────────────────────┘ │ stdio │ HTTP │ HTTP ▼ ▼ ▼ local server mcp.notion.com api.githubcopilot.com/mcp (your machine) (Notion's cloud) (GitHub's cloud) ``` Under the hood every message is JSON-RPC 2.0. The client connects and the two sides negotiate versions and capabilities; the client asks `tools/list` and gets the menu; when the model wants something done, the client sends `tools/call` with arguments, and the result flows back into context. Two transports carry it: **stdio** for servers running as a subprocess on your machine, and **streamable HTTP** for servers at a URL. ## Local vs remote vs in-product | Kind | Where it runs | Setup | Real example | |---|---|---|---| | Remote | The vendor's cloud | Paste a URL, sign in with OAuth | [Notion](https://mcpyet.com/mcp/notion.md), [GitHub](https://mcpyet.com/mcp/github.md) | | Local | Your machine, over stdio | Client launches an npm/PyPI package; you supply API keys | [Playwright](https://mcpyet.com/mcp/playwright.md), [Chrome DevTools](https://mcpyet.com/mcp/chrome-devtools.md) | | In-product | Inside the vendor's own app or CLI | Enable a setting — the product is the server | [GitLab](https://mcpyet.com/mcp/gitlab.md), [Figma](https://mcpyet.com/mcp/figma.md) | Remote is the least friction and keeps credentials with the vendor. Local gives you full control — keys never leave your machine — but the code you run can reach your filesystem, so its provenance matters. In-product servers ride along with software you already trust. ## Official vs community — and why it matters Anyone can write an MCP server for anyone's product, so most products have several. An **official** server is built and maintained by the company behind the product: it tracks their API changes and passes through their security review. A **community** server does the same job unofficially — often well, sometimes abandoned, always unaudited by the vendor. Since a server holds your credentials and sees your data, that difference is the first thing worth knowing about any server you're about to connect. It's the one question every verdict on mcpyet.com answers, with [the evidence attached](https://mcpyet.com/methodology.md). ## How to check whether a product has one 1. **Look it up here.** We track 234 brands — 212 with official servers, 22 with community-only ones: . Each verdict links the vendor's own docs, endpoint or repository. 2. **Check the vendor's docs directly.** Search their documentation for "MCP" — remote servers in particular never show up on GitHub, which is why star-counting directories miss them. 3. **No server yet?** The brand's page here says so plainly, and lists the closest community options if any exist. ## Common questions **What does "MCP server" mean?** MCP stands for Model Context Protocol, an open standard introduced by Anthropic in November 2024. An MCP server is a program that speaks that protocol on behalf of one product — it advertises what the product can do (as tools, resources and prompts) so any compatible AI assistant can use it. **Are MCP servers free?** The servers themselves almost always are — official ones are free software or free hosted endpoints, and community ones are open source. What may cost money is the account behind it: a Notion MCP server is free, but it operates on your Notion workspace, and some vendors gate their API behind paid plans. Some AI clients also require a paid tier for remote connectors. **How do I use an MCP server?** Pick an MCP client (Claude, Cursor, VS Code, ChatGPT and others), then add the server to it. Remote servers need only their URL plus a sign-in; local servers need a launch command in the client's config file. Every brand page on this site carries copy-paste setup for the major clients. **Do I need to know how to code to use one?** Not for remote servers — you paste a URL into your client's settings and sign in. Local servers ask a little more: editing a JSON config file and having Node.js or Python installed. Building a server is a programming task; using one isn't. **What's the difference between MCP and an MCP server?** MCP is the protocol — the shared language, like HTTP. An MCP server is one program that speaks it for one product, like a single website. Clients such as Claude or Cursor are the browsers in this analogy. The protocol explainer: . **Are MCP servers safe to use?** Official servers carry the vendor's own security review and are the safe default. Community servers are unaudited third-party code that you hand real credentials to — research has found SSRF flaws in roughly a third of public servers and no authentication in 41%. Check the maintainer, read the source, and use scoped, revocable keys. --- Source: https://mcpyet.com/what-is-an-mcp-server/ — data refreshed 2026-08-11