Claude Code Review: Anthropic's CLI Coding Agent, Honestly Tested

Pricing and features verified April 2026. Anthropic's lineup moves fast — Opus 4.6, Sonnet 4.6, Haiku 4.5, plugins, skills — so check the docs at claude.com if you're reading this in 2027.

I didn't expect Claude Code to win me over. I'd been using Cursor for a year, the editor view felt right, the tab completion was muscle memory. The idea of dragging my coding work back into a terminal sounded like a regression — like trading a sports car for a stick shift just because someone said it was more authentic.

Then I tried it for a weekend project, and I haven't quite gone back.

Claude Code is Anthropic's command-line AI coding agent. It lives in your terminal, reads and edits files in your repo, runs shell commands, and uses Claude — Opus 4.6 by default — as the brain. That's the surface description. What's underneath is a tool that thinks in entire tasks instead of single completions, and the difference shows up about ten minutes into a real piece of work.

What Claude Code actually is

This is where I struggled at first. Claude Code is not a chatbot, not a copilot in the autocomplete sense, not an IDE. It's an agent that runs in a terminal session and gets full read/write access to whatever directory you point it at — with a permissions prompt the first time it wants to do anything destructive.

You start it with claude, type a request like "audit this repo for missing test coverage and write tests for the three files with the worst coverage," and it goes off and does the thing. Reads files. Writes files. Runs your test suite. Comes back with a report and a diff.

The closest comparison is something like an Aider or a Cursor agents window, but the design choices feel meaningfully different. Claude Code commits hard to the terminal-first idea — it assumes you already have an editor, a git workflow, a test runner. It just glues itself onto the parts of your loop where AI helps.

What works

The model. Claude Opus 4.6 is the strongest model I've used for code that has to actually run, and the Sonnet 4.6 fallback for cheaper turns is genuinely close. I've thrown gnarly Postgres migrations, half-finished React components, and a Python data pipeline at it, and the diffs come back clean enough that I'm reviewing intent more often than syntax.

Plan mode. Press shift+tab and Claude switches into a mode where it lays out exactly what it's about to do — files it'll touch, commands it'll run, the steps in order — and waits for you to approve. The first time I used it, on a refactor that touched seven files, it caught a circular import I would have hit on attempt three. That's the kind of thing the advanced prompt engineering crowd will recognize as a chain-of-thought safety net, except you don't have to engineer it yourself.

Skills, hooks, plugins. Claude Code lets you install reusable "skills" — small markdown files that teach Claude how to do something specific (write a particular kind of test, follow a house style guide, deploy a particular service). Hooks let you run a script before or after Claude touches a file. Plugins bundle skills and MCP servers together. It's all overengineered if you just want to fix a bug, and it's all exactly right if you're trying to make Claude Code part of a team's daily loop. The top Claude skills post has a few I now run on every project.

The honesty about cost. Most agents try to hide token usage. Claude Code shows you, every turn, what you've spent, what model handled what, and whether it switched down to Haiku 4.5 to save money on a small task. I appreciate it.

What doesn't

The terminal is the terminal. If you're a "must see the file open in a real editor" person, you'll feel friction. There are IDE integrations for VS Code and JetBrains that let Claude open a side panel inside your editor, and they help — but the heart of the tool still beats in the CLI. Coming from a GUI-first agent like the ones in Cursor vs Windsurf, that's an adjustment.

It's eager. Without plan mode, Claude will sometimes charge into a task and rewrite five files when you wanted it to fix one. The fix is to use plan mode by default for anything bigger than a typo — but you have to remember, because the default behavior is "go." This is mostly a UX bet I disagree with; I'd rather opt out of caution than into it.

Pricing is awkward to explain to a team. You can run Claude Code on a Claude Pro subscription ($20/month), a Claude Max subscription ($100 or $200/month depending on tier), or with an API key on pay-as-you-go. The Pro plan rate-limits you in ways that get painful on long sessions; the Max plan is generous but a lot of money to justify for one developer; the API is unbounded but you watch the meter spin. Most teams I've talked to land on Max for active developers and Pro for everyone else. Anthropic's pricing page (verified April 2026) is the source of truth — check it before you sign anyone up.

The big-picture caveat: it shines on greenfield and chores, struggles on legacy. On a fresh repo or a contained refactor, Claude Code feels like a senior pairing partner. On a 200k-line monorepo with weird historical decisions, it still needs hand-holding. It's better than Cursor 3 at the smaller end of that spectrum and roughly even at the larger end, in my testing.

Who should use it

Solo developers and small teams who already live in the terminal and want an agent that respects their workflow. People who've been frustrated by GUI-first AI editors trying to take over the screen. Anyone building tools or scripts where Claude can run code, see the output, and iterate.

If you're a designer or PM who codes occasionally, you'll be happier in v0 by Vercel or another GUI-first tool. If you're an enterprise team with strict security around shell access, you'll need to talk to your security people about what Claude Code can touch — the hooks system gives you real controls, but it's not zero-config.

The verdict

Claude Code earned the spot on my dock — or rather, the tab in my terminal. The model quality is the headline, but the design choices around plan mode, skills, and cost transparency are what made me stop reaching for the alternatives. It's not for everyone, and Anthropic still has work to do on the legacy-codebase story, but for the kind of work I do most days, this is now the default.

Frequently Asked Questions

Is Claude Code free?

No, but you can use it with a Claude Pro plan ($20/month, with rate limits) or pay-as-you-go through an API key. The Max plans ($100 or $200/month) give you significantly more headroom for active development work. Verified April 2026 — check Anthropic's pricing page for current numbers.

Does Claude Code work with my IDE?

Yes. There are official integrations for VS Code and JetBrains IDEs that let Claude open a side panel inside your editor, while still running the agent loop in the terminal. The CLI is still the source of truth — the IDE panel is a convenience.

How is Claude Code different from Cursor?

Cursor is an editor with an AI agent built in. Claude Code is an AI agent with no editor — it assumes you already have one. Cursor wins on visual workflow and tab completion; Claude Code wins on planning, skills, and pure model quality.

Can Claude Code run shell commands automatically?

Yes, with your permission. The first time it wants to run something destructive (or write to a file outside its working directory) it asks. You can pre-approve patterns of commands, or run in plan mode where it tells you everything before it does anything.

What model does Claude Code use?

Opus 4.6 by default for the heavy reasoning, with Sonnet 4.6 and Haiku 4.5 used automatically for cheaper subtasks where they're sufficient. You can override the model selection at the command line if you need to.