What is an MCP server, and should you run one?
A small program that gives AI applications tools, resources, and prompts over the Model Context Protocol. Local vs remote, the official registry, and the trust rules before you connect one.
An MCP server is a program that provides context and capabilities to AI applications over the Model Context Protocol: it exposes tools the model can invoke, resources it can read, and prompts it can reuse. “Server” doesn’t mean a machine in a rack: many MCP servers are small local programs your AI app launches as a subprocess and talks to over standard input and output; others run remotely over HTTP and serve many users at once. Connect one to Claude, ChatGPT, or an open-source agent and the model can suddenly read your files, search an issue tracker, or drive a browser.
What one actually looks like
The official reference servers are a good mental picture, because they’re deliberately small: Filesystem (file operations with configurable access controls), Fetch (web content fetching for LLM use), Git (read and search repositories), Memory (a knowledge-graph store), Sequential Thinking, and Time. Each is a modest program that registers a handful of tools and answers JSON-RPC calls. That’s the whole trick: an MCP server is closer to a plugin than to infrastructure. (A caution when reading older articles: many once-official servers for third-party services have been moved to an archived repository, so check what’s current before you depend on one.)
Local and remote are the same protocol
The spec defines two transports, and they map to the two ways servers live. A local server is launched by the client itself and speaks over stdio. When Claude Desktop starts the filesystem server on your machine, that’s this shape, one server per user, dying with the session. A remote server listens over Streamable HTTP (each message an HTTP POST) and typically serves many clients, the way Sentry’s official server runs on Sentry’s own platform. Since the 2026-07-28 spec revision made the protocol stateless, remote servers scale like any ordinary web service: any request can hit any instance behind a load balancer.
Where you find them
The ecosystem’s index is the official MCP Registry at registry.modelcontextprotocol.io, in its own words “like an app store for MCP servers”, run in the open by the project’s contributors, still labeled a preview with its API frozen at v0.1. Third-party directories exist too (mcp.so is a large one), but they’re community catalogs, not the official registry. That’s worth remembering when a listing is your only signal about what a server does.
The part nobody should skip: trust
An MCP server is code you run: the official security guidance is blunt that local servers from untrusted sources can mean arbitrary code execution with your client’s privileges, and it requires clients to warn exactly that. Two more of its rules are worth internalizing: tool descriptions from a server are untrusted unless the server is, and a server must never accept tokens that weren’t issued for it. A server that fetches external content also widens your prompt-injection surface: the model reads whatever the tool returns. The practical posture: treat adding a server like installing software, because it is; prefer official and widely-reviewed servers; and run agents with tool access on a machine that isn’t your personal laptop, where the blast radius is the machine’s, not your life’s.
Your agent, on both sides of the protocol
Agent harnesses are hosts: Claude Code connects to MCP servers by design, and the Claude API’s MCP connector can call remote HTTP servers straight from a Messages request (local stdio servers excluded). OpenClaw goes a step further and sits on both sides: openclaw mcp serve runs the agent as an MCP server, exposing its channel conversations to other MCP clients, while saved definitions under its mcp.servers config let its own runtimes consume outside servers (stdio and streamable-http both supported). Servers a local agent uses run on the agent’s machine, which is one more reason the machine matters more than it first appears, and why they sit alongside plugins and skills in the how-does-my-agent-get-abilities picture.