MCP (Model Context Protocol) is Claude's plug system. Connect it to Linear, GitHub, Sentry, your database, your design tool. Anthropic released MCP in late 2024 and by 2026 it's the standard way LLMs interface with the outside world.
Scope. This lesson is the Claude Code workflow — registering MCP servers in .claude/settings.json, the /mcp slash command, auth patterns, alwaysLoad, and per-server recipes. For the protocol design (the three primitives, JSON-RPC handshake, transports) and how to author an MCP server that any agent can consume, see MCP: Model Context Protocol (protocol + agent architecture).
Learning Objectives
After this lesson, you will be able to:
Register MCP servers in `.claude/settings.json` to give Claude Code access to GitHub, Slack, Postgres, Linear, Figma, and more
Pick the right authentication pattern for any MCP server: OAuth, API key, or dynamic headers
Use `/mcp` slash command to inspect, reconnect, and debug MCP server connections
Apply the 2026 `alwaysLoad` option to make critical MCP tools available without tool-search deferral
Four things are new in 2026: concurrent MCP server connections for faster startup, RFC 9728 token refresh, the alwaysLoad option, and a reserved workspace server name
Build this → Connect the GitHub MCP server to your project; ask Claude to "list open issues with the bug label"; verify it actually queries GitHub via the MCP server
Concretely: the protocol covers tools, resources, and prompts. In Claude Code you register MCP servers in settings, and Claude calls them like any other tool.
Your prompt
↓
Claude Code (orchestrator)
↓
MCP Tool Call (e.g., github.list_issues)
↓
MCP Server (Docker container, npx process, or HTTP endpoint)
↓
External API (GitHub, Slack, Postgres, ...)
/mcp reconnect <name> — restart a disconnected server
/mcp logs <name> — view server stdout/stderr for debugging
The 2026 alwaysLoad Option
By default, Claude Code uses tool-search to discover MCP tools on demand (saves context). For critical tools you always want available, use alwaysLoad:
alwaysLoad: true skips the deferral — the postgres tools are immediately in Claude's tool set. Use sparingly (each always-loaded server adds tokens to every turn).
Claude Code now connects to multiple MCP servers in parallel at startup. Previously sequential, this cut startup time from ~20s to ~3s for projects with 5+ servers.
No config needed — automatic. But verify with /mcp that all your servers connect; if one is slow, it no longer blocks the others.