Claude MCP and Plugins

MCP (Model Context Protocol)

Open standard (Anthropic, 2024), "USB-C for AI": build a server once, any client uses it.

Claude Code (host) ── client 1 ── Server A
                   ── client 2 ── Server B   (one client per server)

A server exposes Tools (callable), Resources (readable), Prompts (templates).
Flow: client fetches tools/list → model picks a tool → client sends tools/call → result added to context.

Transport	Use
HTTP	Remote server (recommended)
stdio	Local process
SSE	Deprecated; v2.1.265+ tries HTTP first and falls back to SSE
WebSocket	Via `add-json` with `"type": "ws"`

Commands

claude mcp add --transport http notion https://mcp.notion.com/mcp
claude mcp add --transport http api https://api.example.com/mcp --header "Authorization: Bearer $TOKEN"
claude mcp add --env KEY=value --transport stdio airtable -- npx -y airtable-mcp-server
claude mcp add-json events '{"type":"ws","url":"wss://example.com/socket"}'
claude mcp list | get <name> | remove <name>
claude mcp login <name> | logout <name> | reset-project-choices
/mcp                       # status, auth, enable/disable
/mcp reconnect all         # retry failed servers

Options go before --; the server command goes after.
Tool search is on by default: tool definitions load on demand via ToolSearch instead of all upfront (ENABLE_TOOL_SEARCH=false disables).

Scopes & Config

Scope	Loads in	Stored in
`local` (default)	This project, private	`~/.claude.json` (under the project path)
`project`	This project, shared via git	`.mcp.json` (needs approval)
`user`	All projects	`~/.claude.json`

Precedence: local > project > user > plugin > claude.ai connectors.

{
  "mcpServers": {
    "api-server": {
      "type": "http",
      "url": "${API_BASE_URL:-https://api.example.com}/mcp",
      "headers": { "Authorization": "Bearer ${API_KEY}" }
    },
    "weather": { "type": "stdio", "command": "npx", "args": ["-y", "my-mcp-server"], "env": {} }
  }
}

${VAR} and ${VAR:-default} expansion supported. Permission rules and approval keys: see Claude Permissions and Claude Settings.

Best Practices

Fewer servers: every tool description costs context.
--scope user for universal, .mcp.json for project.
Trust only what you install (they run code, read files, call APIs).
Pin versions: npx -y package@1.2.3.
Remove unused servers.

Directory: registry.modelcontextprotocol.io. Reference servers: fetch, filesystem, git, memory, time.

Playwright MCP

claude mcp add playwright npx '@playwright/mcp@latest'

Prompt: "Use playwright mcp to open a browser to example.com" or "write an e2e test for login → create task → logout".

Plugins

Bundle skills, agents, hooks, MCP servers, commands, LSP servers, output styles for sharing.

my-plugin/
├── .claude-plugin/plugin.json   # required manifest
├── commands/  agents/  skills/<name>/SKILL.md
├── hooks/hooks.json
├── .mcp.json   .lsp.json   settings.json   bin/   workflows/   monitors/   themes/   output-styles/
└── scripts/  README.md

The manifest is optional; if present only name (kebab-case) is required. Other fields: displayName, version (pins updates), description, author, homepage, repository, license, keywords, defaultEnabled, dependencies, userConfig (prompted values, sensitive option), channels, settings (only agent, subagentStatusLine), and component paths (./-relative, must exist inside the plugin): skills, commands, agents, hooks, mcpServers, lspServers, outputStyles, workflows, experimental.themes|monitors. Extra default dirs: bin/ (on PATH), monitors/, themes/, workflows/, settings.json. claude plugin validate ./my-plugin checks it. A plugin-root CLAUDE.md is not loaded.
Use ${CLAUDE_PLUGIN_ROOT} (changes on update) for bundled files and ${CLAUDE_PLUGIN_DATA} for persistent data. Names starting claude-/anthropic- are reserved.

Marketplaces & Install

Official marketplace claude-plugins-official is added automatically on first interactive start.
/plugin tabs: Discover, Installed, Marketplaces, Errors. /plugin install name@marketplace opens details and a scope choice.
Scopes: user (~/.claude/settings.json), project (.claude/settings.json, teammates still run claude plugin install … --scope project), local (.claude/settings.local.json). Local > project > user.
Add a marketplace: /plugin marketplace add owner/repo (or git URL, path, hosted marketplace.json; #ref pins); one step: /plugin install name --marketplace owner/repo. Also claude plugin install|enable|disable|uninstall|update|marketplace … from the shell.
/reload-plugins applies changes; claude plugin details <name> shows token cost; --plugin-dir loads a local plugin for one session.
Auto-update is on for official marketplaces, off for third-party. Cloud sessions don't load locally installed plugins. Admin controls in Claude Settings.

LSP

Protocol between editors and language servers (Microsoft). Replaces M×N editor/language integrations with M+N. Servers handle parsing, AST, symbols, types, diagnostics, completion, go-to-definition.
Claude Code uses it via the LSP tool. Add a language in a plugin:
command and extensionToLanguage are required; optional args, env, initializationOptions, startupTimeout, requestTimeout (60s default), restartOnCrash, maxRestarts, diagnostics.

{ "go": { "command": "gopls", "args": ["serve"], "extensionToLanguage": { ".go": "go" } } }