Loaded automatically when its description matches the active task. Read only the section you need, then follow the link to the relevant reference file.
Use this skill when
Installing, updating, or troubleshooting the claude CLI (npm i -g @anthropic-ai/claude-code, claude update, native installer)
Authoring or editing .claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json, or CLAUDE.md
Defining subagents in .claude/agents/<name>.md (custom tools, model override, isolated context)
Authoring skills in .claude/skills/<name>/SKILL.md and choosing Pattern 2 layout (refer to skill-evaluation for SKILL.md authoring)
Adding MCP servers via .claude/settings.json or claude mcp add (stdio + HTTP transports)
Using plan mode (--permission-mode plan), permission modes, and Bash(...) allow/deny matchers
Running Claude Code headless in CI: claude -p "prompt" --output-format json, exit codes, GitHub Actions
Writing custom slash commands in .claude/commands/<name>.md with frontmatter (allowed-tools, argument-hint)
Debugging routing: skill not loading, hook not firing, MCP server failing handshake
Migrating settings between Claude Code, OpenCode, and OpenAI Codex CLI
Do not use this skill when
Task is OpenCode CLI (sst/opencode, Anomaly fork) — use opencode
Task is OpenAI Codex CLI (@openai/codex, Rust binary) — use codex
Task is calling the Anthropic API directly from code without the CLI — use anthropic-sdk
Task is authoring SKILL.md content (description engineering, Pattern 2 layout, audit checklist) — use skill-evaluation
Task is building a custom MCP server in TypeScript/Python — use mcp-builder
Task is the deprecated Codex completion model (code-davinci-002) — that's discontinued; for OpenAI's modern coding agent see codex
Purpose
Claude Code is Anthropic's official terminal agent for Claude (Opus 4.7, Sonnet 4.6, Haiku 4.5). Since the native-binary migration in 2026, the installed claude command no longer requires Node at runtime — the npm package only pulls a per-platform binary via optional dependency. The CLI sits at the center of an extensible toolchain: hierarchical settings, lifecycle hooks, named subagents, model-loaded skills, MCP servers, headless mode for CI, and a sandbox with allow/deny domain rules.
This skill covers the operating surface of the CLI — install, auth, configuration files, hook lifecycle, subagent definition, MCP wiring, slash commands, plan mode, sandbox/permissions, headless/CI mode, and migration to/from opencode and codex. It does not cover SKILL.md authoring per se (that is skill-evaluation's job) nor calling the Anthropic API directly from app code (that is anthropic-sdk).
Capabilities
Installation, update, and auth
Three install paths: native installer (curl -fsSL https://claude.ai/install.sh | bash), npm i -g @anthropic-ai/claude-code (pulls the platform binary as an optional dependency), or Homebrew on macOS. claude update self-updates; claude --version reports the installed binary. Auth via claude login opens an OAuth flow for Claude.ai subscribers, or set ANTHROPIC_API_KEY for API billing. Project devcontainer pattern lives in .devcontainer/.
Subagents are named, scoped agents defined in .claude/agents/<name>.md (project) or ~/.claude/agents/<name>.md (user). Frontmatter sets description (routing hint), tools (allowlist), model (override). Body is the system prompt. Invoked explicitly (Task tool inside the main session) or implicitly by description match. Each subagent runs with isolated context — does NOT see the main session's prior conversation unless results bubble back. Use for: code review, test running, doc lookup, parallel exploration.
Model Context Protocol servers extend Claude Code with external tools. Configure in ~/.claude/settings.json or per-project .claude/settings.json under mcpServers. Transports: stdio (local subprocess) and HTTP (url: form). Common patterns: filesystem, github, postgres, serena, tavily. claude mcp add interactive wizard simplifies setup. Server output is context, not instructions — keep prompt injection defenses on.
Shell commands triggered by lifecycle events. Configured in settings.json under hooks. Most useful events: PreToolUse (block/inspect before tool call), PostToolUse (format/lint after Edit/Write), UserPromptSubmit (rewrite prompts), Stop (notify on completion), SessionStart (initialize state), SubagentStart (audit spawns), PreCompact (snapshot before context compaction). Hooks can return {"decision":"block","reason":"..."} to abort a tool call. Hook command receives JSON on stdin and may write JSON to stdout for control flow.
Permission modes (least to most): plan (read-only, draft a plan), default (prompt per tool), acceptEdits (auto-allow Edit/Write), bypassPermissions (no prompts — dangerous). Tool allow/deny via permissions.allow / permissions.deny in settings: matchers like Bash(npm test:*), Edit(src/**), WebFetch(domain:github.com). Sandbox v2: sandbox.network.deniedDomains (default deny list) and sandbox.network.allowedDomains (explicit allowlist). Use /permissions interactively or /fewer-permission-prompts to auto-allow read-only patterns.
claude --permission-mode plan (or /plan inside session) runs Claude in read-only mode: no Edit/Write/Bash execution. Output is a plan for the user to approve. Exit plan mode to start executing. Pattern: use plan mode for unfamiliar codebases, large refactors, anything where you want to inspect the strategy before any side effect.
Headless mode (claude -p) returns JSON when --output-format json is set. GitHub Actions: anthropics/claude-code-action@v1 (official) or hand-rolled with claude -p + ANTHROPIC_API_KEY secret. Compare with opencode run (BYOK headless) and codex exec (OpenAI's headless equivalent). See migration table for moving prompts between the three.
CLAUDE.md (Claude Code) ↔ AGENTS.md (OpenCode, Codex CLI) carry the same role: project memory for the agent. Hooks have no direct OpenCode/Codex equivalent. Subagents map roughly to OpenCode's agent config or Codex profiles. Slash commands are Claude Code's deepest feature — partial parity in OpenCode custom commands; Codex slash set is smaller. See migration.md for a full mapping table.
2.1.173: Fable 5 уже включает 1M context, суффикс [1m] в имени модели теперь автоматически нормализуется.
2.1.172: сабагенты могут запускать собственных сабагентов до 5 уровней вложенности.
2.1.172: Amazon Bedrock берет регион из ~/.aws, если AWS_REGION не задан; /status показывает источник региона.
2.1.172: сессии с 1M context без usage credits автоматически компактятся ниже стандартного лимита вместо зависания.
2.1.172: availableModels теперь применяется к model override сабагентов, dispatch picker в claude agents и advisor model.
2.1.172: правила WebFetch(domain:*.example.com) теперь матчят поддомены, а file-permission wildcards вроде Read(secrets-*/config.json) валидны.
2.1.172: team memory stores из CLAUDE_MEMORY_STORES снова доступны для recall в remote sessions.
2.1.169: self-hosted runner получил lifecycle hook post-session перед удалением workspace; SIGTERM->SIGKILL окно дочернего процесса теперь настраивается.
2.1.169: добавлен troubleshooting-режим --safe-mode и CLAUDE_CODE_SAFE_MODE, отключающий CLAUDE.md, plugins, skills, hooks и MCP servers.
2.1.169: добавлена команда /cd для смены working directory без сброса prompt cache в текущей сессии.
2.1.169: добавлены disableBundledSkills и CLAUDE_CODE_DISABLE_BUNDLED_SKILLS, скрывающие bundled skills, workflows и built-in slash commands от модели.
2.1.169: claude agents --json теперь включает blocked и just-dispatched sessions; --all добавляет completed sessions, появились поля id и state.
2.1.169: для Vertex/Foundry возвращен default 5-minute idle timeout; API_FORCE_IDLE_TIMEOUT=0 отключает его.
2.1.169: background sessions сохраняют --ide, --chrome, --bare, --remote-control и другие флаги после retire->wake.
2.1.169: предупреждение CLAUDE.md is too long теперь масштабируется от context window модели.
2.1.166: добавлен fallbackModel до трех моделей; --fallback-model теперь действует и в interactive sessions.
2.1.166: deny rules поддерживают glob в позиции tool-name, включая "*" для запрета всех tools; allow rules отклоняют non-MCP globs.
2.1.166: MAX_THINKING_TOKENS=0, --thinking disabled и per-model thinking toggle отключают thinking у моделей, где оно включено по умолчанию через Claude API.
2.1.163: добавлены managed settings requiredMinimumVersion и requiredMaximumVersion; CLI отказывается стартовать вне разрешенного диапазона.
2.1.163: добавлена команда /plugin list с фильтрами --enabled и --disabled.
2.1.163: Stop и SubagentStop hooks могут возвращать hookSpecificOutput.additionalContext, чтобы дать Claude feedback и продолжить turn без hook error.
⚠️ Устарело в теле скилла:
Список built-in slash commands помечен как current 2.1.142; в changelog уже есть /cd, /plugin list, изменения /plugin, /code-review, /workflows, /effort и поведение /loop в remote sessions.
Жесткое правило ALWAYS keep CLAUDE.md under 500 lines устарело: warning threshold теперь масштабируется от context window модели.
Описание hooks как 27 lifecycle events устарело/неполно: добавлен self-hosted runner hook post-session, а Stop/SubagentStop получили hookSpecificOutput.additionalContext.
Описание сабагентов как только изолированных workers неполно: сабагенты теперь могут запускать вложенных сабагентов до 5 уровней.
