Skip to main content
SEP 2026 Latest guide updates. Latest: Skill Governance Skill Lifecycle Skills vs Harnesses Changelog →
T12 Intermediate Technical

MCP Servers Overview

The extension protocol and essential servers

PDF
← All cards

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

ServerPrimary use
Context7Official library documentation
Sequential ThinkingStructured multi-step reasoning
GrepaiSemantic search + call graph
SerenaSymbol navigation, refactoring
PlaywrightBrowser automation, E2E tests
GitHubIssues, PRs, repositories

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.

alwaysLoad (v2.1.121)

By default, MCP tools are loaded on demand (via ToolSearch). For servers with 1 to 5 critical tools needed every session, use alwaysLoad: true to make them always available without a prior search call:

{
"mcpServers": {
"my-critical-server": {
"command": "npx",
"args": ["my-mcp-server"],
"alwaysLoad": true
}
}
}

Avoid on servers with many tools (20+): the token cost adds up quickly.

Important limitations

An MCP server sees only the parameters passed at call time, not the conversation history. 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.

Enter your email to read the full card and get the complete PDF bundle.

All content is free and open-source. We just ask for your email.

PDF: