MCP server

#How an MCP server works

A server describes each tool with a name, a description and a JSON Schema for its inputs. The client reads that list when it connects. When the model picks a tool, the client sends a tools/call request, the server runs the tool and returns the result. Servers can also expose resources the app can read and prompts the user can pick.

Servers connect in one of two ways:

• stdio: the app starts the server as a local process and talks to it over standard input and output. No network port opens. See stdio transport.
• Streamable HTTP: the server runs at an HTTP endpoint that clients reach over the network. See Streamable HTTP.

#Why it matters for AI agents

A server is built once and works in any client that speaks MCP. That is why one server can serve Claude Desktop, Claude Code, Cursor and Codex. The server is also where a tool's actions happen. When an agent calls a tool, the server decides what runs, so check which servers you connect.

#Wield and MCP servers

wield-mcp is a local MCP server that runs over stdio. It gives your agent five tools: list_sites, find_recipe, run, request_recipe and status. run executes a recipe in your own browser, so the requests carry your normal browser session. Passwords and cookies stay on your computer, and results do not pass through wield.run. See the tool reference.

#Related terms

• MCP client: the side of the connection inside the AI app.
• Tool calling: how the model asks for a tool to run.
• Model Context Protocol: the protocol that defines both sides.

#FAQ

#Can an MCP server run on my own computer?

Yes. A local MCP server runs as a process that your AI app starts, and it talks to the app over stdio. No network port is opened. wield-mcp works this way: your AI app starts it on your computer when you use it.

#What is the difference between an MCP server and a REST API?

A REST API is a set of endpoints that a service offers over HTTP. A caller must already know the endpoints and their formats. An MCP server lists its tools with names and input schemas, and a client reads that list at runtime. An MCP server often calls one or more REST APIs inside its tools.