MCP Servers Overview

What is MCP?

MCP (Model Context Protocol) is the open standard that allows Claude Code to connect to external tools via a JSON-RPC 2.0 protocol. Each MCP server exposes additional tools, named according to the mcp__<server>__<tool> convention. Claude uses them exactly like its native tools, with the same permissions system.

Basic architecture: Claude Code (client) communicates with the MCP server via stdio or HTTP. The server starts on first use and remains active for the entire session.

Most widely used servers

Configuration in settings.json

MCP servers are declared in ~/.claude/settings.json (global) or .claude/settings.json (project).

{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp"],
      "env": {}
    },
    "github": {
      "command": "npx",
      "args": ["@github/mcp-server"],
      "env": {
        "GITHUB_TOKEN": "${GITHUB_TOKEN}"
      }
    }
  }
}

Transport: local (stdio) vs remote (HTTP)

stdio (local): the server runs as a child process on the machine. No network exposure, ideal for local tools like grepai (Ollama) or Serena.

HTTP/SSE (remote): the server is accessible via a URL. Used for hosted services like Figma MCP (https://mcp.figma.com/mcp) or shared team servers.

{
  "mcpServers": {
    "figma": {
      "transport": "http",
      "url": "https://mcp.figma.com/mcp"
    }
  }
}

MCP Apps (SEP-1865)

Stable extension since January 2026, co-developed by Anthropic and OpenAI. It allows MCP servers to return interactive interfaces (HTML/JS) in addition to classic text responses, rendered in a sandboxed iframe on the client side.

Use cases: data dashboards, complex forms, real-time visualizations directly within the conversation.

Important limitations

An MCP server does not access the conversation history — it only sees the parameters passed at call time. Each call is independent unless the server itself implements a cache. It also cannot modify Claude’s system prompt or bypass the permissions system.

Token cost: each loaded server adds approximately 2K tokens of overhead per session (tool definitions). Load only what is necessary for the current project.