Claude Code — Agents, Skills, Plugins

March 03, 2026 AI,, Claude,, agents,, skills,, plugins,, automation

Overview

Claude Code extends beyond the base model through a layered extensibility system: agents (specialized subagents), skills (packaged expertise and workflows), and plugins (bundles of all of the above). Together they let you tailor Claude to your project and team.

Skills

Skills are markdown files (SKILL.md) that teach Claude how to perform specific tasks. Claude can invoke them automatically when context suggests they’re relevant, or you can invoke them manually via /skill-name.

Aspect Details
Structure Directory with SKILL.md, optional reference.md, examples.md, scripts/
Locations User: ~/.claude/skills/ — Project: .claude/skills/ — Plugin: bundled in plugin
Invocation disable-model-invocation: true → user-only (deploy, commit, send)
  user-invocable: false → Claude-only (background knowledge)
  Default → both can invoke

When to create skills:

  • Frequently repeated prompts or workflows
  • Project-specific tasks with arguments
  • Applying templates or scripts
  • Workflows that should run in isolation (context: fork)

Agents (Subagents)

Agents are specialized subagents Claude invokes for specific domains. They run in parallel, have their own system prompts, and can use tools.

Use Case Agent Type
Large codebase (>500 files) code-reviewer — parallel code review
Auth/payments code security-reviewer — security audits
API project api-documenter — OpenAPI generation
Performance critical performance-analyzer — bottleneck detection
Frontend heavy ui-reviewer — accessibility review
Needs more tests test-writer — test generation

Location: .claude/agents/<name>.md

Skills vs Subagents

Aspect Skill Subagent
What it is Instructions the current agent follows Separate agent instance
Context Same as main agent Isolated / focused
Execution Main agent applies the skill Runs autonomously with its own tools

When to use skills: Single task, linear workflow, context sharing is useful.

When to use subagents:

  • Focus — Main agent accumulates too much context and drifts. Subagents stay on one job.
  • Fresh perspective — Subagent reviews “cold” without main agent’s assumptions; catches what the main agent misses.
  • Parallelism — Run security + performance + code review at once.
  • Isolation — Different tool permissions, smaller context, failure containment.

Skills extend the current agent. Subagents delegate to a separate agent. Use subagents when you need independence, parallelism, or focus — not just different instructions.

Plugins

Plugins are self-contained directories that bundle multiple extension types. One install gives you skills, agents, hooks, MCP servers, and optionally LSP servers.

Plugin Structure

plugin-name/
├── .claude-plugin/plugin.json   # required manifest
├── commands/                    # slash commands
├── agents/                      # subagent definitions
├── skills/                      # agent skills
├── hooks/                       # event handlers (hooks.json)
├── .mcp.json                    # MCP server config
└── scripts/

Plugin Components

Component Purpose
Skills /name shortcuts, model-invoked capabilities
Agents Specialized subagents for specific tasks
Commands Custom slash commands
Hooks Event handlers (PreToolUse, PostToolUse, etc.)
MCP servers External tool integrations (DBs, APIs, browsers)
LSP servers Language intelligence for code navigation

Hook Events (examples)

  • PreToolUse / PostToolUse — before/after tool execution
  • PostToolUseFailure — after tool execution fails
  • SessionStart / SessionEnd
  • UserPromptSubmit
  • SubagentStart / SubagentStop

Automation Types at a Glance

Type Best For
Hooks Automatic actions on tool events (format on save, lint, block edits)
Subagents Specialized reviewers/analyzers that run in parallel
Skills Packaged expertise, workflows, repeatable tasks
Plugins Collections of skills (and agents, hooks, MCP) — install as a unit
MCP Servers External tool integrations (databases, APIs, docs)

Discovery & Install

  • Community registry: claude-plugins.dev — CLI for discovering and installing plugins
  • Official marketplace: Search in Claude Code’s /plugin Discover tab
  • MCP servers: claude mcp add <server-name> (e.g. context7, playwright, github)

Plugin Distribution (Without Official Marketplace)

To distribute plugins without publishing to the official Anthropic marketplace:

1. Custom marketplace — Host a marketplace.json that catalogs your plugins. Users add it with:

/plugin marketplace add ./my-marketplace              # local path
/plugin marketplace add owner/repo                    # GitHub
/plugin marketplace add https://gitlab.com/company/plugins.git
/plugin marketplace add https://example.com/marketplace.json

Then install plugins from that marketplace via /plugin install.

2. Git repo — Put the plugin in a public or private repo. If the repo root has a marketplace.json that references the plugin, users add the repo as a marketplace and install from there.

3. Local / direct share — For internal use: share the plugin directory (zip, clone, copy). Users run claude --plugin-dir ./path-to-plugin or add a local marketplace pointing to that path.

4. Community registry — Publish to claude-plugins.dev instead of the official marketplace.

Key Insight

Skills teach what to do. Agents specialize who does it. Plugins package both for reuse. Hooks and MCP servers handle when and with what tools — the glue that makes automation automatic.