How to Use Claude Code in 2026 — From Install to Your First Commit

The URL still says 2025, but the content was fully rewritten for August 2026. The address stays put so existing links and search results don't break. Last updated: 2026-08-25

Summary: How do you use Claude Code in 2026?

• Claude Code is no longer a terminal-only CLI. Terminal, VS Code, JetBrains, the desktop app, web, mobile, Slack, and Chrome all share the same settings and the same CLAUDE.md.
• Install with the native installer, not npm install -g.
• The models are Opus 5, Sonnet 5, Fable 5, and Haiku 4.5. The first three carry a 1M-token context window.
• Reasoning depth is set by effort (low through max), not a token budget. budget_tokens returns a 400 on current models.
• Since 2026-08-14, auto mode is the default permission mode on Pro, Max, and Team.

Most Claude Code tutorials were written in 2025. The install command has changed, the model names have changed, and the way you set reasoning depth has been removed. Even the opening line — "a CLI tool you run in your terminal" — is no longer true. So before the install steps, start with where you can actually run it.

1. Where Claude Code runs

New features land in the terminal CLI first. But there are now eight doors in: the VS Code and Cursor extension (inline diffs, plan review), the JetBrains plugin, the desktop app (visual diffs, parallel sessions, scheduled work), the web at claude.ai/code with nothing installed locally, the mobile apps, an @Claude mention in Slack, and Chrome.

The count matters less than the fact that they all read the same CLAUDE.md and the same MCP configuration. Sessions move between them too: claude --teleport pulls a web session into your terminal, and /desktop throws a terminal session into the desktop app.

2. Installing

curl -fsSL https://claude.ai/install.sh | bash   # macOS, Linux, WSL
irm https://claude.ai/install.ps1 | iex          # Windows PowerShell

Package managers work too: brew install --cask claude-code and winget install Anthropic.ClaudeCode. The Homebrew claude-code cask tracks a stable channel that lags about a week behind; claude-code@latest follows the newest release. Only the native installer updates itself in the background.

On Windows, Git for Windows is no longer required. Since 2026 week 18 (v2.1.120), Claude Code falls back to PowerShell as its shell tool when Git Bash isn't present. After installing, cd into a project and run claude. If something looks broken, /doctor inspects your whole setup and repairs what it can.

3. Your first thirty minutes

Run /init. It scans the project and drafts a CLAUDE.md, the file loaded automatically at the start of every session. Resist the urge to make it big. The official guidance is blunt: for each line, ask whether deleting it would make Claude get something wrong, and if not, delete it. A long CLAUDE.md gets half-ignored. Placement and merge order are covered in the CLAUDE.md setup guide.

Press Shift+Tab for plan mode. Claude proposes an approach without touching files. Fixing a plan is far cheaper than reverting code. If you want a more formal methodology, see spec-driven development.

Give it something it can check. This is what separates a good session from a frustrating one. "Fix the login bug" is worse than "fix it until npm test -- auth passes." Claude needs a way to grade itself — tests, a build, a screenshot comparison. To pin that condition for the whole session, use /goal. Why the verification loop matters is laid out in harness engineering for AI agents.

Commit. Ask it to split the changes into logical commits and it will read the diff and write the messages. If you don't like where things went, press Esc twice or use /rewind.

Strung together: in plan mode, say "add duplicate-email checking to the signup form; finish with npm test -- signup passing," read and correct the plan, Shift+Tab out to let it run, and commit once the tests are green.

4. Models and effort

Don't append dates to model IDs — claude-opus-5 is the complete ID. Opus 5 needs Claude Code v2.1.219+, Sonnet 5 needs v2.1.197+, Fable 5 needs v2.1.170+, and Fable 5 is not a default; pick it with /model fable.

Reasoning depth comes from effort: five levels from low to max, defaulting to high, changed with /effort xhigh. Two things 2025 posts routinely get wrong live here. Setting a numeric thinking budget via budget_tokens is gone and returns a 400 on current models. And while ultrathink still works as a keyword, think hard is not one — it's just text.

5. Permissions and context

Since 2026-08-14, interactive sessions on Pro, Max, and Team default to auto mode, where a separate classifier model reviews what Claude is about to do and blocks only the risky parts. There are six permission modes, cycled with Shift+Tab. That's why the old advice to disable permission checking wholesale no longer applies — the fully bypassing mode has been demoted to something you use inside an isolated container or VM, not in everyday work.

The context window is still the fundamental constraint on quality. Switching tasks calls for /clear; continuing a long one calls for /compact <instruction>; /rewind restores a checkpoint; /btw asks a side question that never enters the transcript; /context shows what's currently loaded. An unscoped "go investigate this codebase" is the classic way to burn the window.

6. What to delete from a 2025 tutorial

7. Where to go next

Claude Code isn't in the Free plan; it starts at Pro. Max raises the limits, Team comes in Standard and Premium tiers, and Enterprise is per-seat. Prices are displayed differently by billing currency, so check the official pricing page. /usage breaks your remaining limit down by skill, subagent, plugin, and MCP server, and the peak-hour allocation change is covered in Claude Code usage limits.

If you want a GUI on macOS for running several jobs at once, see running Claude Code in parallel. If you want external events pushed into a live session, see Claude Code Channels — Channels is a notification path into a session, not a parallelism tool.

Sources

• Anthropic Claude Code documentation