Plugin & Skill Extension System
Claude Code offers two extension mechanisms: plugins for structured feature extensions, and skills for reusable prompt workflows.
Plugin System
Plugins are versioned installable packages contributing commands, skills, MCP servers, hooks, and settings. Loaded via pluginLoader.ts, merged with session and enterprise sources.
Plugin Lifecycle
Plugin commands merge into getCommands(), MCP servers into getClaudeCodeMcpConfigs(), hooks via loadPluginHooks.ts.
Skill System
Skills are model-invokable prompt workflows, implemented as Command objects with type: 'prompt'. Not a separate runtime -- they're loaded into the command registry and executed via SkillTool.
Skill Types
| Type | Source | Format |
|---|---|---|
| Bundled | TypeScript modules | Compiled into binary, optional disk file extraction |
| Disk | skills/ directory | SKILL.md (Markdown + frontmatter) |
| Plugin | Plugin skills/ paths | Markdown files |
| MCP | MCP prompts | Synthesized commands (feature-gated) |
SkillTool Execution
SkillTool finds the target skill from available commands, determines execution method (fork vs inline), and runs it.
Command Priority Chain
Bundled skills > Built-in plugin skills > Disk skills > Workflows
> Plugin commands > Plugin skills > Built-in slash commands
Key Source Files
| File | Responsibility |
|---|---|
src/utils/plugins/pluginLoader.ts | Plugin loading and merging |
src/skills/bundledSkills.ts | Bundled skill definitions |
src/skills/loadSkillsDir.ts | Disk skill loading |
src/tools/SkillTool/SkillTool.ts | Skill execution engine |
src/commands.ts | Command registry (includes skills) |
Next
Go to 12-api-streaming.md to learn about API calls and streaming.
Hands-on Experiment
This chapter has a corresponding Python experiment:
Covers: SKILL.md parsing, plugin lifecycle, command priority chain