Claude Code

Claude Code is Anthropic's agentic coding tool that lives in the terminal, reads your repo, runs your shell, edits files, and drives git on your behalf. It works as an interactive REPL, a scriptable headless command for CI, a long-running background agent, or a session you steer from claude.ai or your phone. It plugs into VS Code, Cursor, Windsurf, JetBrains, GitHub Actions, Chrome, and MCP servers, so the same agent shows up wherever you already work.

The pitch is straightforward: instead of autocomplete, you get an assistant that can plan a change, touch the files, run the tests, open a PR, and tell you what it did.

What It Is

Claude Code is a CLI you install with one curl or PowerShell line and launch by typing claude in any repo. It then has read, write, edit, search, and Bash tools, plus whatever MCP servers and plugins you've connected. The docs describe a default loop where Sonnet handles most coding work, Opus is reserved for planning and harder reasoning, and Haiku covers cheap mechanical tasks. Models, effort levels, and permission modes are all swappable mid-session.

Core Capabilities

• Agentic codebase search — Claude explores files on its own with grep, glob, and reads to answer architecture questions and trace dependencies without you hand-selecting context.
• Direct edits with diffs — Reads, writes, and edits files in place, showing proposed diffs you can approve, edit, or reject.
• Bash execution — Runs your build, tests, linters, formatters, and any other CLI tool already on your machine.
• Git and PR automation — Stages, commits, branches, rebases, and opens GitHub pull requests via the gh CLI.
• Plan mode and checkpointing — Drafts a plan before large changes and lets you rewind code and conversation to any earlier checkpoint with /rewind.
• MCP, plugins, and skills — Connects to external tools (Jira, Slack, Notion, databases, custom services) through MCP; plugins extend behavior, and skills are markdown prompt files Claude can invoke automatically.

Parallel, Background, and Scheduled Work

/batch decomposes a large change into 5–30 independent units and runs each one in its own git worktree as a background subagent that opens a PR when it finishes. /background detaches the current session so it keeps working while you free up your terminal, and claude agents is the management view for monitoring, attaching to, and respawning those sessions.

For recurring or watchful work, /loop runs a prompt repeatedly while the session is open (on an interval like 5m or self-paced) and /schedule creates routines that execute on Anthropic-managed cloud infrastructure. /autofix-pr spawns a web session that watches a PR and pushes fixes when CI fails or reviewers leave comments. The same agent also runs from your browser via Claude Code on the web, with /teleport to pull a web session back into your terminal and Remote Control to drive a local session from claude.ai or the Claude mobile app.

Command Cheat Sheet

The CLI exposes dozens of slash commands. The docs group them roughly by where they fit in a session:

Launch-time flags cover automation: -p for print/headless mode, --output-format json, --max-turns, --max-budget-usd, --permission-mode, --allowedTools, --mcp-config, --bg, and --remote.

Usage and Limits to Plan Around

Anthropic publishes a session-and-week model for subscription usage. The support docs describe a rolling 5-hour session window plus separate weekly limits, with a stricter weekly cap on Opus than on other models. The Usage settings page surfaces both bars. Hitting a limit prints a "resets at time" message; you can switch to a lighter model or, on eligible plans, enable extra usage.

Every turn re-sends the whole conversation, your CLAUDE.md, and any files Claude has read, which is where most token spend comes from. The docs recommend five habits: /clear between tasks, match the model to the job, reference files by path instead of pasting, keep CLAUDE.md lean (under ~200 lines), and ask for a plan before large changes. The /model opusplan shortcut bakes the "plan with Opus, execute with Sonnet" pattern in. For API users (Console, Bedrock, Vertex, Foundry), there is no hard cap; /cost shows running session spend.

Update: Recent Additions

Recent additions called out in the official docs include the agent view and background sessions, /batch with worktree-based parallel work, /loop and /schedule for time-based work, Remote Control, Claude Code on the web with /teleport, /ultrareview for deep multi-agent review in a cloud sandbox, plugins and channels (research preview), and /extra-usage for continuing past subscription limits.