[proxy] code.claude.com← back | site home | direct (HTTPS) ↗ | proxy home | ◑ dark◐ light

Extend Claude Code - Claude Code Docs

Claude Code combines a model that reasons about your code with built-in tools for file operations, search, execution, and web access. The built-in tools cover most coding tasks. This guide covers the extension layer: features you add to customize what Claude knows, connect it to external services, and automate workflows.

New to Claude Code? Start with CLAUDE.md for project conventions. Add other extensions as you need them.

Overview

Extensions plug into different parts of the agentic loop:

Skills are the most flexible extension. A skill is a markdown file containing knowledge, workflows, or instructions. You can invoke skills with a command like /deploy, or Claude can load them automatically when relevant. Skills can run in your current conversation or in an isolated context via subagents.

Match features to your goal

Features range from always-on context that Claude sees every session, to on-demand capabilities you or Claude can invoke, to background automation that runs on specific events. The table below shows what’s available and when each one makes sense.

FeatureWhat it doesWhen to use itExample
CLAUDE.mdPersistent context loaded every conversationProject conventions, “always do X” rules”Use pnpm, not npm. Run tests before committing.”
SkillInstructions, knowledge, and workflows Claude can useReusable content, reference docs, repeatable tasks/review runs your code review checklist; API docs skill with endpoint patterns
SubagentIsolated execution context that returns summarized resultsContext isolation, parallel tasks, specialized workersResearch task that reads many files but returns only key findings
Agent teamsCoordinate multiple independent Claude Code sessionsParallel research, new feature development, debugging with competing hypothesesSpawn reviewers to check security, performance, and tests simultaneously
MCPConnect to external servicesExternal data or actionsQuery your database, post to Slack, control a browser
HookDeterministic script that runs on eventsPredictable automation, no LLM involvedRun ESLint after every file edit

Plugins are the packaging layer. A plugin bundles skills, hooks, subagents, and MCP servers into a single installable unit. Plugin skills are namespaced (like /my-plugin:review) so multiple plugins can coexist. Use plugins when you want to reuse the same setup across multiple repositories or distribute to others via a marketplace.

Compare similar features

Some features can seem similar. Here’s how to tell them apart.

Skills and subagents solve different problems:

  • Skills are reusable content you can load into any context
  • Subagents are isolated workers that run separately from your main conversation
AspectSkillSubagent
What it isReusable instructions, knowledge, or workflowsIsolated worker with its own context
Key benefitShare content across contextsContext isolation. Work happens separately, only summary returns
Best forReference material, invocable workflowsTasks that read many files, parallel work, specialized workers

Skills can be reference or action. Reference skills provide knowledge Claude uses throughout your session (like your API style guide). Action skills tell Claude to do something specific (like /deploy that runs your deployment workflow).Use a subagent when you need context isolation or when your context window is getting full. The subagent might read dozens of files or run extensive searches, but your main conversation only receives a summary. Since subagent work doesn’t consume your main context, this is also useful when you don’t need the intermediate work to remain visible. Custom subagents can have their own instructions and can preload skills.They can combine. A subagent can preload specific skills (skills: field). A skill can run in isolated context using context: fork. See Skills for details.

Both store instructions, but they load differently and serve different purposes.

AspectCLAUDE.mdSkill
LoadsEvery session, automaticallyOn demand
Can include filesYes, with @path importsYes, with @path imports
Can trigger workflowsNoYes, with /<name>
Best for”Always do X” rulesReference material, invocable workflows

Put it in CLAUDE.md if Claude should always know it: coding conventions, build commands, project structure, “never do X” rules.Put it in a skill if it’s reference material Claude needs sometimes (API docs, style guides) or a workflow you trigger with /<name> (deploy, review, release).Rule of thumb: Keep CLAUDE.md under 200 lines. If it’s growing, move reference content to skills or split into .claude/rules/ files.

All three store instructions, but they load differently:

AspectCLAUDE.md.claude/rules/Skill
LoadsEvery sessionEvery session, or when matching files are openedOn demand, when invoked or relevant
ScopeWhole projectCan be scoped to file pathsTask-specific
Best forCore conventions and build commandsLanguage-specific or directory-specific guidelinesReference material, repeatable workflows

Use CLAUDE.md for instructions every session needs: build commands, test conventions, project architecture.Use rules to keep CLAUDE.md focused. Rules with paths frontmatter only load when Claude works with matching files, saving context.Use skills for content Claude only needs sometimes, like API documentation or a deployment checklist you trigger with /<name>.

Both parallelize work, but they’re architecturally different:

  • Subagents run inside your session and report results back to your main context
  • Agent teams are independent Claude Code sessions that communicate with each other
AspectSubagentAgent team
ContextOwn context window; results return to the callerOwn context window; fully independent
CommunicationReports results back to the main agent onlyTeammates message each other directly
CoordinationMain agent manages all workShared task list with self-coordination
Best forFocused tasks where only the result mattersComplex work requiring discussion and collaboration
Token costLower: results summarized back to main contextHigher: each teammate is a separate Claude instance

Use a subagent when you need a quick, focused worker: research a question, verify a claim, review a file. The subagent does the work and returns a summary. Your main conversation stays clean.Use an agent team when teammates need to share findings, challenge each other, and coordinate independently. Agent teams are best for research with competing hypotheses, parallel code review, and new feature development where each teammate owns a separate piece.Transition point: If you’re running parallel subagents but hitting context limits, or if your subagents need to communicate with each other, agent teams are the natural next step.

MCP connects Claude to external services. Skills extend what Claude knows, including how to use those services effectively.

AspectMCPSkill
What it isProtocol for connecting to external servicesKnowledge, workflows, and reference material
ProvidesTools and data accessKnowledge, workflows, reference material
ExamplesSlack integration, database queries, browser controlCode review checklist, deploy workflow, API style guide

These solve different problems and work well together:MCP gives Claude the ability to interact with external systems. Without MCP, Claude can’t query your database or post to Slack.Skills give Claude knowledge about how to use those tools effectively, plus workflows you can trigger with /<name>. A skill might include your team’s database schema and query patterns, or a /post-to-slack workflow with your team’s message formatting rules.Example: An MCP server connects Claude to your database. A skill teaches Claude your data model, common query patterns, and which tables to use for different tasks.

Understand how features layer

Features can be defined at multiple levels: user-wide, per-project, via plugins, or through managed policies. You can also nest CLAUDE.md files in subdirectories or place skills in specific packages of a monorepo. When the same feature exists at multiple levels, here’s how they layer:

Combine features

Each extension solves a different problem: CLAUDE.md handles always-on context, skills handle on-demand knowledge and workflows, MCP handles external connections, subagents handle isolation, and hooks handle automation. Real setups combine them based on your workflow. For example, you might use CLAUDE.md for project conventions, a skill for your deployment workflow, MCP to connect to your database, and a hook to run linting after every edit. Each feature handles what it’s best at.

PatternHow it worksExample
Skill + MCPMCP provides the connection; a skill teaches Claude how to use it wellMCP connects to your database, a skill documents your schema and query patterns
Skill + SubagentA skill spawns subagents for parallel work/review skill kicks off security, performance, and style subagents that work in isolated context
CLAUDE.md + SkillsCLAUDE.md holds always-on rules; skills hold reference material loaded on demandCLAUDE.md says “follow our API conventions,” a skill contains the full API style guide
Hook + MCPA hook triggers external actions through MCPPost-edit hook sends a Slack notification when Claude modifies critical files

Understand context costs

Every feature you add consumes some of Claude’s context. Too much can fill up your context window, but it can also add noise that makes Claude less effective; skills may not trigger correctly, or Claude may lose track of your conventions. Understanding these trade-offs helps you build an effective setup.

Context cost by feature

Each feature has a different loading strategy and context cost:

FeatureWhen it loadsWhat loadsContext cost
CLAUDE.mdSession startFull contentEvery request
SkillsSession start + when usedDescriptions at start, full content when usedLow (descriptions every request)*
MCP serversSession startAll tool definitions and schemasEvery request
SubagentsWhen spawnedFresh context with specified skillsIsolated from main session
HooksOn triggerNothing (runs externally)Zero, unless hook returns additional context

*By default, skill descriptions load at session start so Claude can decide when to use them. Set disable-model-invocation: true in a skill’s frontmatter to hide it from Claude entirely until you invoke it manually. This reduces context cost to zero for skills you only trigger yourself.

Understand how features load

Each feature loads at different points in your session. The tabs below explain when each one loads and what goes into context.

When: Session startWhat loads: Full content of all CLAUDE.md files (managed, user, and project levels).Inheritance: Claude reads CLAUDE.md files from your working directory up to the root, and discovers nested ones in subdirectories as it accesses those files. See How CLAUDE.md files load for details.

Skills are extra capabilities in Claude’s toolkit. They can be reference material (like an API style guide) or invocable workflows you trigger with /<name> (like /deploy). Claude Code ships with bundled skills like /simplify, /batch, and /debug that work out of the box. You can also create your own. Claude uses skills when appropriate, or you can invoke one directly.When: Depends on the skill’s configuration. By default, descriptions load at session start and full content loads when used. For user-only skills (disable-model-invocation: true), nothing loads until you invoke them.What loads: For model-invocable skills, Claude sees names and descriptions in every request. When you invoke a skill with /<name> or Claude loads it automatically, the full content loads into your conversation.How Claude chooses skills: Claude matches your task against skill descriptions to decide which are relevant. If descriptions are vague or overlap, Claude may load the wrong skill or miss one that would help. To tell Claude to use a specific skill, invoke it with /<name>. Skills with disable-model-invocation: true are invisible to Claude until you invoke them.Context cost: Low until used. User-only skills have zero cost until invoked.In subagents: Skills work differently in subagents. Instead of on-demand loading, skills passed to a subagent are fully preloaded into its context at launch. Subagents don’t inherit skills from the main session; you must specify them explicitly.

When: Session start.What loads: All tool definitions and JSON schemas from connected servers.Context cost: Tool search (enabled by default) loads MCP tools up to 10% of context and defers the rest until needed.Reliability note: MCP connections can fail silently mid-session. If a server disconnects, its tools disappear without warning. Claude may try to use a tool that no longer exists. If you notice Claude failing to use an MCP tool it previously could access, check the connection with /mcp.

When: On demand, when you or Claude spawns one for a task.What loads: Fresh, isolated context containing:

  • The system prompt (shared with parent for cache efficiency)
  • Full content of skills listed in the agent’s skills: field
  • CLAUDE.md and git status (inherited from parent)
  • Whatever context the lead agent passes in the prompt

Context cost: Isolated from main session. Subagents don’t inherit your conversation history or invoked skills.

When: On trigger. Hooks fire at specific lifecycle events like tool execution, session boundaries, prompt submission, permission requests, and compaction. See Hooks for the full list.What loads: Nothing by default. Hooks run as external scripts.Context cost: Zero, unless the hook returns output that gets added as messages to your conversation.

Learn more

Each feature has its own guide with setup instructions, examples, and configuration options.