Beyond chat, it has extension points that change how it works: hooks that run your own commands at lifecycle events, subagents with their own context, skills (which now absorb custom slash commands), plan mode for read-only planning, MCP servers for external tools, and non-interactive print mode (claude -p) for scripts and CI. This brief covers what each does, when it is worth turning on, and how to enable it, for developers already using Claude Code on a Pro or Max plan or an API key.
*Field guide compiled 2026-10-23 from the official documentation, pricing pages and changelogs linked below; every line links to the page that states it.*
What it does
- Hooks. User-defined shell commands, HTTP endpoints, MCP tool calls, LLM prompts or subagents that run automatically at lifecycle events such as PreToolUse, PostToolUse, UserPromptSubmit, SessionStart and Stop, receiving JSON context and optionally returning a decision. Use them for things that must always happen (formatting, blocking dangerous commands, notifications) rather than relying on the model. How: Define under "hooks" in ~/.claude/settings.json (all projects), .claude/settings.json (shared with the repo) or .claude/settings.local.json; nest event, matcher group (e.g. only the Bash tool) and handler. Browse with the /hooks menu. source
- Subagents. Specialized assistants that each run in their own context window with a custom system prompt, tool access and permissions, and return only a summary. Use one when a side task would flood the main conversation with search results, logs or file contents; Claude delegates automatically when a task matches a subagent's description. How: Create Markdown files with frontmatter in .claude/agents/ (project, commit to git) or ~/.claude/agents/ (personal), or pass JSON with --agents for one session; you can also ask Claude to create one or invoke one explicitly. source
- Skills. A SKILL.md file of instructions that Claude adds to its toolkit and loads only when used, so long procedures cost little context until needed. Claude can use a skill automatically when relevant, or you invoke it with /skill-name. Use one when you keep pasting the same checklist or procedure. How: Create .claude/skills/name/SKILL.md with name and description frontmatter. Set disable-model-invocation: true for side-effect workflows like /deploy so only you can trigger them, or user-invocable: false for background knowledge only Claude loads. source
- Custom commands. Custom slash commands have been merged into skills: a file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work the same way; existing .claude/commands/ files keep working. Built-in commands such as /help and /compact are listed in the commands reference. How: Type / to see available commands or filter them; text after the command name becomes its arguments ($ARGUMENTS in the skill body). source
- Plan mode. Tells Claude to research and propose changes without making them: it reads files, runs shell commands to explore and writes a plan, but edits stay blocked until you approve the plan. Use it before larger or riskier changes, or to explore an unfamiliar codebase. How: Press Shift+Tab until the status bar shows plan mode on, prefix a single prompt with /plan, or start with claude --permission-mode plan. Make it the default by setting defaultMode to plan in .claude/settings.json. source
- MCP. Model Context Protocol servers connect Claude Code to external tools, databases and APIs (issue trackers, docs, monitoring). Remote HTTP, legacy SSE, local stdio and WebSocket servers are supported, and tool search loads tools on demand by default. How: claude mcp add --transport http name url (e.g. claude mcp add --transport http notion https://mcp.notion.com/mcp); add --scope project to write .mcp.json for the team, or use local (default) or user scope stored in ~/.claude.json. source
- non-interactive mode. Runs Claude Code programmatically from scripts and CI/CD via the Agent SDK CLI, returning text, JSON with result and session ID, or streamed newline-delimited JSON. Use it for pipelines, build scripts, commit generation and piping data through Claude. How: claude -p "prompt" (or --print) with options such as --allowedTools "Read,Edit,Bash", --output-format json|stream-json, --continue, and --bare for reproducible runs that skip auto-discovery. source
Modes
- default (Manual). Only reads run without asking; you review every action. source
- acceptEdits. Reads, file edits and common filesystem commands run without prompts. source
- plan. Reads plus classifier-approved commands when auto mode is available; edits wait for plan approval. source
- auto. Everything runs, with background safety checks; for long tasks. source
- dontAsk. Only file reads in working directories and pre-approved tools; anything that would prompt is denied. For locked-down CI and scripts. source
- bypassPermissions. Everything runs without checks; isolated containers and VMs only. source
Pricing and limits
Recent changes
- 2026-10-22: v2.1.309: fixed sandboxed Bash commands running shell startup files outside the sandbox, and a disabled MCP server being reconnected by a later permission prompt in print and SDK sessions. source
- 2026-10-21: v2.1.307: added the claudecode.safetyevent OpenTelemetry event for tool calls stopped or prompted by a permission rule, auto mode, a hook or the sandbox; fixed MCP servers showing needs authentication in claude -p on macOS with a locked keychain. source
- 2026-10-21: v2.1.306: fixed claude -p inside cloud sessions reporting Not logged in, and spurious prompt is too long errors in sessions with many MCP tools. source
- 2026-10-19: v2.1.304: added /auto-mode-setup, a guided setup that teaches auto mode about your environment. source
- 2026-10-16: v2.1.302: added CLAUDECODEAUTOMODEMODEL to choose the client-side auto mode classifier model, and an Idle compaction toggle in /config. source
Traps
- Without --bare, a claude -p session runs the hooks in a project's .claude/settings.json and connects the servers in its .mcp.json, even in a folder you have never trusted, with no workspace trust dialog or per-server approval prompt. source
- Bare mode (--bare) skips hooks, skills, subagents, plugins, MCP servers and CLAUDE.md, and does not use your subscription login, so ANTHROPICAPIKEY must be set. source
- Subagents send their own requests, which count toward the same usage limits as the main conversation; by default spawning fails once 20 subagents are running (CLAUDECODEMAXCONCURRENTSUBAGENTS changes it). source
- Cloud sessions don't read your local ~/.claude/settings.json, so user-level hooks there will not fire in them. source
- The SSE transport for MCP is deprecated; use HTTP servers where available. source