Claude Code
terminal CLIVS CodeDesktopwebJetBrains
Claude Code has the broadest explicit extension taxonomy in a single `.claude` ecosystem. CLAUDE.md supports multiple layers: managed policy (OS-level, cannot be excluded), project, user, local, and ancestor directory layers with on-demand subdirectory loading. The @import syntax allows CLAUDE.md files to reference other files recursively. Rules (`.claude/rules/*.md`) provide path-scoped instructions. Skills are the most flexible extension — reference skills provide knowledge, action skills trigger workflows, and they can run in isolated context via subagents. Agent Teams are experimental: multiple independent Claude Code sessions that coordinate via shared tasks and peer-to-peer messaging. Hooks support 4 handler types across 27 lifecycle events. Plugins include plugin-only components (LSP servers, monitors, channels) with marketplace distribution, dependency management, and managed restrictions.
CLAUDE.md and Rules
InstructionsVendor Terms CLAUDE.md, CLAUDE.local.md, .claude/CLAUDE.md, rules, .claude/rules/*.md, @import syntax, auto memory, MEMORY.md
Interfaces terminal CLI, VS Code, Desktop, web, JetBrains
Scopes
Organization / EnterpriseUser / HomeProject / Repo RootSubdirectory / Folder
Availability current (explicit)
Trust Model Context files, not enforced configuration; Anthropic explicitly says compliance is not guaranteed for vague or conflicting instructions; managed policy CLAUDE.md deployed by IT cannot be excluded
Notes CLAUDE.md walks up directory tree; subdirectory files load on demand; rules can be path-scoped; managed policy CLAUDE.md at OS-level paths cannot be excluded; @path import syntax with 5-hop recursive depth; AGENTS.md supported via @import; claudeMdExcludes for monorepo filtering; auto memory is a complementary per-working-tree system
Skills
PromptsVendor Terms skills, SKILL.md, commands, custom commands, slash commands, .claude/skills/<name>/SKILL.md, .claude/commands/*.md, bundled skills, Agent Skills standard
Interfaces terminal CLI, VS Code, Desktop, web
Scopes
Project / Repo RootUser / HomeOrganization / Enterprise
Availability current (explicit)
Trust Model Reusable knowledge and workflows; auto-loaded when relevant or invoked via /name; skills can pre-approve tools via allowed-tools frontmatter; enterprise managed settings can disable shell execution in skills
Notes Skills are the primary prompt/workflow mechanism; legacy .claude/commands/ still works. Dual invocation: user via /name and model-invoked automatically. Follows Agent Skills open standard (agentskills.io). Bundled skills: /batch, /debug, /simplify, /loop, /claude-api, /fewer-permission-prompts. MCP prompts surface as /mcp__<server>__<prompt>. Enterprise deployment via managed settings.
Skills (knowledge and workflows)
SkillsVendor Terms skills, SKILL.md, reference skills, action skills, Agent Skills standard, bundled skills
Interfaces terminal CLI, VS Code, Desktop, web
Scopes
Project / Repo RootUser / HomeOrganization / Enterprise
Availability current (explicit)
Trust Model Reusable knowledge and workflows; auto-loaded when relevant or invoked via /name; bundled skills like /simplify, /batch, /debug included; enterprise deployment via managed settings
Notes Reference skills provide knowledge, action skills trigger workflows; can run in isolated context via subagents. Follows Agent Skills open standard. Rich frontmatter: model, effort, context:fork, agent, hooks, paths, allowed-tools. Dynamic context injection via shell preprocessing. Auto-discovered from nested directories (monorepo support). Skill content persists with auto-compaction budget (5K tokens each, 25K combined).
MCP Servers
MCP & ToolsVendor Terms MCP servers, .mcp.json, local-scoped MCP servers, project-scoped MCP servers, user-scoped MCP servers, managed-mcp.json, MCP Tool Search, MCP resources, MCP prompts, MCP elicitation, channels, plugin-provided MCP servers, claude.ai connectors
Interfaces terminal CLI, VS Code, Desktop, web, JetBrains
Scopes
User / HomeProject / Repo RootOrganization / Enterprise
Availability current (explicit)
Trust Model Project-scoped servers require explicit approval; managed-mcp.json provides enterprise lockdown; allowlist/denylist policies via managed settings; plugin MCP servers inherit plugin trust
Notes Three installation scopes: local (default, per-project private), project (shared via .mcp.json), and user (cross-project). Tool Search enabled by default defers tool loading. Supports HTTP, SSE (deprecated), and stdio transports. OAuth 2.0 auth for remote servers. Plugin-provided MCP servers auto-start. MCP resources, prompts, and elicitation supported. Channels enable push messages. Claude Code can serve as an MCP server. Enterprise managed-mcp.json and allowlist/denylist policies available.
Subagents
AgentsVendor Terms subagents, custom subagents, managed subagents, built-in subagents, plugin subagents, project subagents
Interfaces terminal CLI, VS Code
Scopes
User / HomeProject / Repo RootOrganization / EnterpriseCloud / Web Session
Availability current (explicit)
Trust Model Isolated workers with their own context, prompt, tools, and optionally model; only summary returns to main conversation; permission mode inheritance from parent; plugin subagents cannot use hooks/mcpServers/permissionMode
Notes Five priority levels: managed > CLI flag > project > user > plugin. Built-in subagents: Explore, Plan, general-purpose, statusline-setup, Claude Code Guide. Supports isolation worktrees, persistent memory, skills preloading, scoped MCP servers, hooks in frontmatter, effort levels, background mode, @-mention invocation, and subagent resumption.
Agent Teams
AgentsVendor Terms agent teams, teammates, team lead, shared task list, mailbox, peer-to-peer messaging
Interfaces terminal CLI
Scopes
Project / Repo RootUser / Home
Availability experimental (explicit)
Trust Model Teammates inherit lead's permission settings; can change individual modes after spawning but not at spawn time; if lead uses --dangerously-skip-permissions all teammates do too
Notes Experimental and disabled by default; requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Multiple independent Claude Code sessions with shared tasks and peer-to-peer messaging. Two display modes: in-process and split-pane (tmux/iTerm2). Task dependencies, quality gate hooks (TeammateIdle, TaskCreated, TaskCompleted), plan approval workflow.
Hooks
HooksVendor Terms hooks, shell commands, HTTP endpoints, LLM prompts, agent hooks, async hooks, prompt hooks, MCP tool hooks, asyncRewake, matcher, decision control, CLAUDE_ENV_FILE
Interfaces terminal CLI, VS Code
Scopes
User / HomeProject / Repo RootOrganization / Enterprise
Availability current (explicit)
Trust Model Deterministic control layer supporting shell commands, HTTP endpoints, LLM prompts, and agent subagents; hooks run with full user permissions; can intercept lifecycle events and tool use; organization admins can restrict to managed hooks only
Notes 4 handler types (command, http, prompt, agent) across 27 lifecycle events; async hooks run in background; hooks defined in settings files, plugins, skills, and agent frontmatter; PreToolUse hooks enforce policy even in bypass mode; allowManagedHooksOnly lets org admins restrict to managed hooks
Plugins and Marketplaces
Plugins & DistributionVendor Terms plugins, plugin.json, marketplace.json, .claude-plugin, LSP servers, monitors, marketplace, plugin hints, plugin dependencies, output styles, channels, userConfig, bin
Interfaces terminal CLI, VS Code, Desktop, web
Scopes
Project / Repo RootUser / HomeOrganization / Enterprise
Availability current (explicit)
Trust Model Plugins execute arbitrary code with user privileges; managed marketplace restrictions allow org lockdown; plugin caching isolates installs; path traversal outside plugin root blocked; official marketplace requires submission review
Notes Includes plugin-only components (LSP servers, monitors) not available as standalone files; plugins namespace skills as /plugin-name:skill. Official Anthropic marketplace with submission review. npm packages as plugin source. Plugin dependency version constraints with semver. Plugin hints protocol for CLI integration. Managed marketplace restrictions (strictKnownMarketplaces).
Settings and Managed Settings
Settings & PolicyVendor Terms settings.json, settings.local.json, managed settings, managed-settings.json, managed-settings.d/, server-managed settings, ~/.claude.json, managed preferences (plist), registry policies
Interfaces terminal CLI, VS Code, Desktop, web
Scopes
Organization / EnterpriseUser / HomeProject / Repo Root
Availability current (explicit)
Trust Model Formal policy/config layer with documented precedence; managed settings delivered by server policy or OS-level device management cannot be overridden; controls permissions, hooks, environment variables, model defaults
Notes Four-tier scope: Managed > CLI args > Local > Project > User. 50+ settings keys covering permissions, sandbox, hooks, env, plugins, auto-mode, and more. Server-managed settings via Claude.ai admin console for Teams/Enterprise. MDM delivery via macOS plist, Windows registry, or file-based managed-settings.json with drop-in directory. Managed-only settings cannot be overridden.
Output Styles
Settings & PolicyVendor Terms output styles, system prompt variants, custom output styles, outputStyle setting, output-styles/ directory
Interfaces terminal CLI, VS Code, Desktop, web
Scopes
Project / Repo RootUser / Home
Availability current (explicit)
Trust Model Stronger than ordinary context guidance because it changes the system prompt layer; intended to style behavior
Notes Directly modifies Claude Code's system prompt; three built-in styles (Default, Explanatory, Learning); custom styles as Markdown files with YAML frontmatter; plugins can distribute output styles