Claude Code

The agentic coding tool — a CLI you install locally, a redesigned desktop app, an IDE extension, and a browser interface at claude.ai/code that runs tasks on Anthropic-managed cloud VMs.

What it is

Claude Code is where Claude takes multi-step actions against code: reading files, running commands, editing, running tests, opening PRs. It has four forms that share the same underlying engine.

  • CLI — install on your machine, run claude in any repo, work in your terminal.
  • Desktop app — redesigned 2026-04-14 around parallel sessions: a session sidebar across repos, drag-and-drop panes (terminal / preview / diff / chat), an in-app file editor and diff viewer, SSH on macOS (already supported on Linux), and one Git worktree per session.
  • IDE extension — VS Code and JetBrains, surfacing the same engine in your editor.
  • Web — claude.ai/code runs the agent in an Anthropic-managed cloud VM that clones your GitHub repo and opens PRs. Use --teleport from the CLI to hand off a cloud session into your terminal, or claude --bg to run a session in the background and reattach later via /resume or claude agents (a single-screen view of every running, blocked, and finished session; accepts --add-dir, --settings, --mcp-config, --plugin-dir, --model, --effort). Inside claude agents, prefix any shell command with ! to spin it up as a detachable background job — equivalent to claude --bg --exec '<command>' from your shell.

Claude Code is where Skills, MCP servers, Plugins, Hooks, and slash commands all install.

When to use it

  • Multi-step coding against files in a real repo.
  • Anything you’d want to leave running unattended — long migrations, refactors, test runs.
  • Async work via claude.ai/code — kick off jobs from the iOS app or web and check back later.
  • Setting up Routines (/schedule) for recurring tasks.
  • Letting Claude grind on a measurable goal across many turns — /goal "<completion condition>" (v2.1.139+, May 2026) keeps Claude working until a separate evaluator model confirms the condition holds (e.g., npm test exits 0).
  • Tackling a migration or audit too large for one session — include the keyword ultracode in your prompt (renamed from workflow in v2.1.160 — the bare word workflow no longer fires), or set /effort ultracode to auto-orchestrate the whole session, to fan the job out across up to 1,000 parallel subagents (research preview, requires v2.1.154+; see Dynamic Workflows on the surfaces page).
  • Reviewing a GitHub PR without installing the GitHub App. From a checked-out PR branch, run /code-review to print findings in the terminal, /code-review --comment to post them as inline PR comments (≥ 80-confidence, deduped), or /code-review --fix to apply them to the working tree. Originally absorbed /simplify in v2.1.146 (2026-05-21); since v2.1.154 (2026-05-28), /simplify is back as a separate cleanup-only review (reuse, simplification, efficiency, altitude — no bug-hunting). If you want bug-hunting plus apply-fixes, use /code-review --fix.

How to install / enable

  • macOS/Linux: native installer.

    curl -fsSL https://claude.ai/install.sh | bash
claude
  • Windows PowerShell: irm https://claude.ai/install.ps1 | iex, or winget install Anthropic.ClaudeCode (use winget upgrade Anthropic.ClaudeCode to bump an existing install).
  • Homebrew: brew install --cask claude-code (stable) or claude-code@latest (newest).
  • Linux distros: apt, dnf, apk packages also available.
  • npm install (npm install -g @anthropic-ai/claude-code) still works but Anthropic now prints a yellow “npm installation is deprecated” banner. Use the native installer for new setups.
  • Web: https://claude.ai/code (Pro, Max, Team, or premium Enterprise seats).
  • Pick a model with /model (or --model at launch). Opus 4.8 is the default. Claude Fable 5 was selectable from v2.1.170 (2026-06-09) but is currently unavailable: on 2026-06-12 a U.S. government export-control directive forced Anthropic to disable Fable 5 (and Mythos 5) for all customers worldwide. Anthropic is contesting the order and says it’s working to restore access, with no timeline. All other models, including Opus 4.8, are unaffected, so the picker works normally — Fable 5 just won’t appear until access is restored.

Common pitfalls

  • Claude Code on the web runs in a sandboxed VM with network restrictions; local-only MCP servers won’t work there.
  • Skills/plugins/MCP installed in ~/.claude/ (user scope) are global; .claude/ in a repo (project scope) is per-repo.
  • Background sessions accumulate — list with claude agents and clean up. Pin one with Ctrl+T in that view if you want it kept alive when idle and restarted in place on updates; un-pinned sessions are shed first under memory pressure.
  • The planned Agent SDK billing split was paused. Anthropic had announced that claude -p (headless mode), Claude Code GitHub Actions, and any third-party harness using the Agent SDK would move to a separate credit pool on 2026-06-15. On that date Anthropic shelved the change: for now, claude -p and the Agent SDK still draw from your normal subscription, same as interactive claude in the terminal. Anthropic says it will give notice before any future change. See Claude API.

See also

Sources

  • Claude Code product landing — Anthropic; verified 2026-05-29 (this run, via search summary) — canonical install command (curl -fsSL https://claude.ai/install.sh | bash) and OS-specific options, including winget install Anthropic.ClaudeCode.
  • Set up Claude Code — Anthropic docs; verified 2026-05-26 (this run).
  • Claude Code on the web — Anthropic docs; verified 2026-05-19.
  • Redesigning Claude Code on desktop for parallel agents — Anthropic blog; published 2026-04-14 — session sidebar, drag-and-drop panes, SSH on macOS, per-session Git worktrees.
  • Claude Code changelog (v2.1.139–v2.1.162) — claude agents config flags, /code-review (absorbed /simplify in v2.1.146 on 2026-05-21; gained --fix in v2.1.152 on 2026-05-27; gained --comment for inline PR comments in v2.1.161, 2026-06-02). In v2.1.154 (2026-05-28) /simplify returned as a distinct cleanup-only command (reuse / simplification / efficiency / altitude — no bug-hunting), so it is no longer an alias for /code-review --fix. Also: claude --bg, claude --bg --exec '<command>', and the ! prefix inside claude agents to spin background jobs; /resume (also covers background sessions as of v2.1.161); Ctrl+T pinned background sessions; /goal (v2.1.139, 2026-05-12); /reload-skills + disallowed-tools frontmatter (v2.1.152); Opus 4.8 + Dynamic Workflows + /effort ultracode (v2.1.154, 2026-05-28); verified 2026-06-05 (this run).
  • Code review — Anthropic docs; verified 2026-06-03 (this run) — /code-review, --fix, --comment flag semantics; 80-confidence threshold; parallel review subagents.
  • Keep Claude working toward a goal — Anthropic docs; verified 2026-05-27 — /goal completion-condition loop with separate evaluator model, 4,000-char limit, requires v2.1.139+.
  • Introducing dynamic workflows in Claude Code — Anthropic blog; published 2026-05-28 — research preview, 1,000-agent cap, 16 concurrent, plan availability (Max/Team on by default, Enterprise admin-gated, Pro toggle in /config).
  • Orchestrate subagents at scale with dynamic workflows — Anthropic docs; verified 2026-06-06 (this run) — single-task keyword renamed workflow → ultracode in v2.1.160; /effort ultracode session setting (xhigh + auto-orchestration); v2.1.154 requirement; CLAUDE_CODE_DISABLE_WORKFLOWS env var.
  • Install Claude Code the Right Way in 2026 — secondary; npm-deprecation banner and native-installer migration path; verified 2026-05-25.
  • Use the Claude Agent SDK with your Claude plan — Anthropic help center; verified 2026-06-20 (this run) — opening notice confirms the planned Agent SDK / claude -p billing split is paused as of 2026-06-15; usage still draws from subscription limits.
  • Anthropic puts Claude agents on a meter across its subscriptions — InfoWorld; published 2026-05-14 — the (now-paused) plan to move claude -p and the Agent SDK onto a separate credit pool on 2026-06-15.
  • Claude Fable 5 and Claude Mythos 5 — Anthropic news; published 2026-06-09 — Fable 5 selectable in Claude Code v2.1.170; Opus 4.8 remains default.
  • Statement on the US government directive to suspend access to Fable 5 and Mythos 5 — Anthropic news; published 2026-06-12; verified 2026-06-20 (this run) — export-control directive received 5:21pm ET 2026-06-12; Fable 5 and Mythos 5 disabled for all customers worldwide; all other models (incl. Opus 4.8) unaffected; no restoration timeline.