Claude Code tutorial: plugins, skills, and the top MCP servers to install today

Claude Code is Anthropic's official CLI coding agent. Out of the box it can read, edit, and run code on your machine. With plugins, skills, and MCP servers wired in, it turns into something closer to a full engineering environment that understands your repo, your tools, and your external systems — and brings them all into one conversation.

This is the walkthrough we give to every customer's engineering team on day one. No fluff. Copy the commands, try them, and you'll have a production-grade Claude Code setup in under an hour.

The three primitives: plugins, skills, MCP servers

Before you install anything, understand the difference — it's the thing most teams get wrong in week one.

• Plugins are bundles of commands, skills, agents, and hooks that ship together. When you install impeccable, you get a set of /impeccable:* slash commands, a suite of skills, and the agents that power them. One install, many tools.
• Skills are model-invoked capabilities. You don't call them directly — Claude Code decides to use them when the conversation matches their trigger. A well-written skill is like a team member who quietly jumps in with exactly the right checklist when a task fits their specialty.
• MCP servers (Model Context Protocol) are long-lived external tools that expose their own functions to Claude. Unlike a plugin, an MCP server can run anywhere — locally, in a container, on a remote host — and it exposes typed capabilities Claude can call during any conversation. Think of MCP as the bus between Claude and everything outside your filesystem.

Rule of thumb: use a plugin when you want pre-wired workflows, a skill when you want Claude to quietly do the right thing on its own, and an MCP server when you want Claude to reach into an external system it otherwise can't see.

Installing your first plugin

Plugins live in a marketplace. The happy path is two commands:

claude plugin marketplace add <marketplace-url>
claude plugin install <plugin-name>

After install, run /reload-plugins inside Claude Code to activate without restarting. New slash commands, skills, and agents show up immediately.

You can verify an install worked by typing / and scanning the command list. If the plugin ships commands they'll appear with their namespace prefix (e.g., /impeccable:critique).

Installing an MCP server

MCP servers are added through your settings. The fastest way is through Claude Code's config:

claude mcp add <server-name> -- <command to run the server>

For example, to wire in a Postgres MCP that lets Claude query a local database:

claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres postgresql://localhost/mydb

Restart Claude Code. The server's tools are now available to the model. Ask it to query the database and it will call the MCP server under the hood.

Top 10 MCP servers worth installing this week

This is our actual install list — the ones we add to every new workspace by default. Every one of these pays for itself on the first conversation.

1. filesystem — official. Scoped filesystem access for a chosen directory. Essential when Claude Code is running outside your repo.
2. github — official. Read/search/create issues, PRs, files, gists. Turns the GitHub REST API into something the model can navigate natively.
3. postgres — official. Read-only SQL queries against a live Postgres. Lets Claude answer questions about your actual production data during debugging.
4. sqlite — official. Same idea for SQLite. Great for local dev and embedded data stores.
5. puppeteer — official. Headless browser control. Claude can load a page, click, screenshot, and scrape — useful for E2E verification and visual QA.
6. brave-search — web search that doesn't require an OpenAI key. Clean results, attribution, predictable rate limits.
7. slack — read channels, post messages, search history. The difference between "report back to the team" being an interruption vs. a background action.
8. linear — triage tickets, create issues from code review, link commits to work items. Straightens out the engineering → project-management loop.
9. sentry — pull live error data. When an agent is debugging a regression, having Sentry in-context beats pasting stack traces.
10. context7 — fetches current, versioned documentation for libraries and frameworks. Stops Claude from guessing an API that changed six months ago.

Honorable mentions: memory, git, fetch, google-drive, notion, stripe. Install based on your stack.

Top 10 Claude Code plugins worth installing this week

Plugins are higher-level than MCP servers — they bundle skills, agents, and commands you'd otherwise have to assemble yourself.

1. superpowers — a suite of skills for planning, writing plans, executing them, systematic debugging, test-driven development, parallel subagent dispatch, and code review. The most common "first plugin" we install for a new engineer.
2. impeccable — design review, critique, polish, animate, layout, and a dozen more design-focused skills. If you ship any frontend at all, this is non-optional. (We used it to critique and fix every page on the site you're reading this on.)
3. code-review — pull-request-style review for any diff. Runs independent checks, identifies risk, and explains why before suggesting changes.
4. searchfit-seo — a full SEO toolkit: audits, content strategy, keyword clustering, schema markup, on-page optimization, internal linking analysis, AI visibility tracking.
5. frontend-design — opinionated UI generator that avoids the generic AI-slop aesthetic. Ships creative, production-grade frontend code instead of templated card grids.
6. ui-ux-pro-max — a broader UX intelligence library: design patterns, palettes, font pairings, chart types, UX guidelines, stack-specific recommendations.
7. claude-api — skill for building, debugging, and optimizing Claude API / Anthropic SDK apps. Handles prompt caching patterns automatically.
8. update-config — configures the Claude Code harness itself through settings.json. Needed for hooks, automated behaviors, and workflow rules.
9. keybindings-help — the fastest way to rebind keys, add chord shortcuts, or customize ~/.claude/keybindings.json. Saves an hour on day one.
10. loop and schedule — run prompts or commands on a recurring interval or a cron schedule. Turns one-off Claude sessions into always-on automations.

A realistic day one setup

Here's the exact sequence we run on a new machine. Ten minutes end to end.

# 1. Install Claude Code (if you haven't already)
# Follow the official docs at claude.com/claude-code

# 2. Wire in the MCP servers you'll actually use
claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /Users/you/code
claude mcp add github -- npx -y @modelcontextprotocol/server-github
claude mcp add context7 -- npx -y @context7/mcp-server

# 3. Install the plugins
claude plugin marketplace add <your-marketplace-url>
claude plugin install superpowers
claude plugin install impeccable
claude plugin install code-review

# 4. Reload without restarting
# Inside Claude Code:
/reload-plugins

# 5. Verify
# Type "/" and confirm the new commands show up.
# Ask Claude "what MCP servers are connected?" and you should get your list.

That's it. You now have a Claude Code that can reach into your filesystem, your GitHub, live documentation, and has design critique, planning, debugging, and code review wired in as skills.

What to use first

Three experiments that prove the setup works:

• Critique a real page: open any file with markup and run /impeccable:critique. You'll get a design health score, priority issues, and suggested fixes.
• Write a plan from a one-line spec: use the superpowers:writing-plans skill. Hand it a sentence of what you want to build; get a structured plan back.
• Review a diff: commit some changes, don't push, run /code-review:code-review. It will walk your actual diff, flag risks, and propose fixes.

If all three work, you're set. If any fail, the most common fix is a /reload-plugins followed by a restart.

Things to avoid

• Don't install every plugin on day one. Pick 3–5 that match what you're actually doing this week. Adding all ten at once turns your command palette into noise.
• Don't wire in MCP servers you don't have credentials for. A half-configured server is worse than no server — Claude will try to use it and fail silently.
• Don't skip the marketplace verification. Only install plugins from marketplaces you trust. Plugins run in the same process as Claude Code and can touch your filesystem.
• Don't treat skills as commands. Skills are model-invoked — you describe what you want, Claude picks the skill. If you're typing /use-this-skill explicitly, you probably want a plugin command instead.

Want us to set up a Claude Code environment for your entire engineering team in one afternoon? That's Path 1 — Train your team — we run it as a hands-on workshop and your engineers leave with a working, customized install.