Claude Power-Ups · Reference

Claude Code ships with several distinct extension mechanisms. They overlap in confusing ways. This page cuts through it: what each thing is, how it differs, and when to reach for it.

Mental Model · Layers

Tools

Native functions Claude can call: Read, Write, Bash, Grep, WebSearch…

Foundation

MCP Servers

Add more tools by connecting external servers via the Model Context Protocol

Extend tools

Connectors

Pre-built MCP servers for named apps (Slack, Gmail, Linear, GitHub…)

MCP for apps

Skills

Saved slash-command workflows that orchestrate existing tools

Saved workflows

Plugins

Installable bundles of multiple skills, packaged and distributed

Skill bundles

Agents / Subagents

Isolated Claude sessions spawned to do parallel or expensive work

Parallel work

Hooks

Shell commands that fire automatically on Claude events (start, stop, tool calls)

Automation

🔧

Tools

The base layer — everything else builds on these

Foundation

Native functions Claude can call directly. These ship with Claude Code and cannot be removed. Every other extension mechanism either adds more tools or orchestrates existing ones.

Think of as: the standard library. You don't install them. They're always there.

Read — open a file

Bash — run a shell command

Grep — search code

WebSearch — search the web

Agent — spawn a subagent

Always active

🔌

MCP Servers

Model Context Protocol — adds new tools

Extend tools

External servers Claude connects to that expose additional tools. MCP is a standard protocol, so any developer can build and publish an MCP server. Once connected, Claude sees those tools alongside its native ones.

Think of as: USB peripherals. Plug one in, Claude gains new capabilities. Unplug it, they're gone.

computer-use — mouse and keyboard control

claude-in-chrome — browser DOM access

preview_* — dev server control

Any custom server you build

Use when: need Claude to touch something it can't natively

🔗

Connectors

Pre-packaged MCP for popular services

MCP for apps

Connectors are just MCP servers, but pre-built and maintained for specific named products. You install them instead of building from scratch. They handle auth, API quirks, and schema so you don't have to.

Think of as: MCP servers with a brand name on the box. Slack connector = MCP server for Slack.

Slack — read/write messages and channels

Gmail — read email, draft replies

Linear — create and update tickets

GitHub — PRs, issues, code search

Use when: want Claude inside a specific named app

⚡

Skills

Slash commands — saved, repeatable workflows

Saved workflows

Markdown files containing instructions Claude follows when you type /skillname. They encode domain knowledge, step sequences, and behavior rules for tasks you run repeatedly. Claude reads the skill file and executes accordingly.

Think of as: macros. Write the instructions once, invoke with a slash command forever.

/edit-dashboard — safe edit protocol with checks

/pillars-management — bulk data operations

/recap — harvest insights into memory

/caveman — activate terse response mode

Use when: task is repeatable and has known steps

📦

Plugins

Installable bundles of multiple skills

Skill bundles

Plugins package several related skills together for distribution. You install a plugin and get all its slash commands at once. Anthropic and third parties can publish plugins; you can build your own.

Think of as: an app that installs multiple slash commands. One install, many capabilities.

caveman plugin — adds /caveman, /caveman-review, /caveman-commit

cowork plugin — adds session, memory, and planning skills

Custom domain plugin — bundle all your team's workflows

Use when: someone ships a workflow package worth installing

🤖

Agents & Subagents

Isolated Claude sessions for parallel or isolated work

Parallel work

Claude can spawn separate Claude sessions (subagents) to run tasks in parallel or in isolation. Each subagent gets its own context, tool set, and instructions. Results come back as a single message. The spawning Claude is the "agent"; what it spawns is a "subagent."

Think of as: forking threads. Main Claude dispatches work; subagents do it in parallel; results merge back.

Run QA audit while main session continues

Explore 3 different files simultaneously

Isolate expensive research from main context

Specialized types: Explore, Plan, general-purpose

Use when: work is parallelizable or needs isolation

🧠

Memory

Two distinct systems people constantly conflate

Persistence

CLAUDE.md — always-loaded instructions. Every session reads these automatically. Two levels: global (~/.claude/CLAUDE.md) and project (./CLAUDE.md). Use for rules, preferences, and constraints that always apply.

Memory files — saved facts recalled on demand. Four types: user (who you are), feedback (how to work with you), project (current context), reference (where to find things). Claude loads these when relevant, not always.

Key distinction: CLAUDE.md = rules that always fire. Memory files = context recalled when relevant. Don't put rules in memory files — Claude may not load them.

CLAUDE.md: "Never use em dashes. Always move files to archived/, never delete."

Memory: "User is a data scientist. Prefers backend analogies for frontend concepts."

Memory: "Feedback: don't add trailing summaries — user can read the diff."

Rules → CLAUDE.md · Facts → Memory files

🛡️

Permissions & Settings

settings.json — controls what Claude can do without asking

Config

Claude prompts you before every tool call by default. settings.json lets you pre-approve categories of actions so Claude can work autonomously. This is also where Hooks live.

Think of as: sudo rules. Define what Claude is allowed to do unattended vs. what requires your explicit approval each time.

Allowlist Bash(git *) — no prompt for any git command

Allowlist Read(*) — Claude reads files silently

Denylist Bash(rm *) — always block, no override

Hooks defined here fire automatically on events

Two scopes: user-level (~/.claude/) and project-level (.claude/)

Use when: tired of approving the same tool calls repeatedly

🎚️

Model Tiers

Opus · Sonnet · Haiku — mix per task

Performance

Three tiers with different capability and cost tradeoffs. The main session uses one model, but each spawned agent can run on a different one. Use expensive models where reasoning matters; cheap models for mechanical lookups.

Key insight: you can mix models mid-session by specifying the model when spawning an agent. Run Opus for planning, Haiku for file search, Sonnet for implementation.

Haiku — fastest, cheapest. File searches, simple lookups, mechanical edits.

Sonnet — balanced default. Most coding, writing, moderate reasoning.

Opus — most capable, most expensive. Complex reasoning, architecture, ambiguous problems.

Agents accept a model param: "haiku" / "sonnet" / "opus"

Match model tier to task complexity — don't default everything to Opus

⚙️

Hooks

Shell commands triggered by Claude events

Automation

Hooks are shell commands wired to lifecycle events in Claude Code. They run automatically without Claude having to remember. Configured in settings.json. The output feeds back to Claude as context.

Think of as: git hooks, but for Claude actions. Event fires, shell runs, Claude sees the result.

UserPromptSubmit — fires before every Claude reply

Stop — fires when Claude finishes a response

PreToolCall — fires before any tool runs

Common use: inject active mode reminders, auto-log sessions

Use when: behavior should be automatic, not remembered

When to Use What

I want to…	Reach for	Not
Give Claude access to a specific app	Connector (if one exists)	Tools
Give Claude access to a custom system	MCP Server	Connector
Automate a repeatable workflow	Skill	Agent
Install a set of workflows someone built	Plugin	Skill (one-by-one)
Run heavy work without burning main context	Agent	Inline in main session
Run independent tasks at the same time	Agents (parallel)	Sequential tool calls
Enforce behavior without Claude remembering	Hook	Memory files
Control mouse, keyboard, desktop apps	MCP (computer-use)	Connector
Interact with a web app in a browser	MCP (claude-in-chrome)	MCP (computer-use)
Set rules/preferences that always apply	CLAUDE.md	Memory files
Save facts about user/project for later	Memory files	CLAUDE.md
Stop getting prompted for routine tool calls	settings.json allowlist	Asking Claude to remember
Run expensive reasoning task	Opus agent	Default Sonnet for everything
Run fast/cheap lookup or search	Haiku agent	Opus (overkill)