Claude Code: AI Agent That Edits Your Project Autonomously
Claude Code reads your project, changes files, runs tests and fixes errors. How it works, four things that make the biggest difference, and real costs.
Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.
👁 116 reads
Claude Code is an AI coding agent that works in your terminal. It doesn’t suggest lines in your editor - it takes a task, reads your project files, makes changes, runs commands, checks results, and fixes itself when something breaks. The difference between this and editor assistants is like the difference between suggesting and executing, and that’s what determines whether a tool saves you hours or just minutes.
How Claude Code Works
You launch it in your project directory and describe what needs to happen. The agent figures out the code, proposes a plan, makes changes, and shows them for approval. It has shell access, so it can run tests, check what failed, and fix it without asking you for error details. It also works with version control, which is crucial here: when working with an agent, change history stops being a formality and becomes a safety net.
Beyond the terminal, it’s available in a desktop app, in popular development environments, and in the browser. It’s the same agent, just in different packaging.
Four Things That Make the Biggest Difference
- Project rules file. A short document describing conventions, how to run tests, and things that shouldn’t be touched saves more time than any prompt trick.
- Small tasks. One change, one approval. When a task is described in three sentences, it’s hard to pinpoint which step failed.
- End condition. “Change this so that test passes” is verifiable. “Fix this” is not.
- Reading changes. Code can be correct and simultaneously inconsistent with the rest of the project. After a month of such approvals, you have two ways of doing the same thing.
What Claude Code Costs
Access comes either through a subscription or pay-as-you-go billing through the API. With the first, you pay a fixed amount and stay within usage limits; with the second, you pay for what you actually process, and on a large project it can surprise you. The rule is simple: the more context the agent needs to load, the more expensive it gets, so keeping your project organized translates to your bill. I collect current rates for all versions in the article how much Claude costs.
How to Start in Ten Minutes
You need a Claude account with a Pro or Max plan (Claude Code doesn’t work on the free plan) or an API key billed by tokens. Installation is one terminal command: npm install -g @anthropic-ai/claude-code, then in your project directory you type claude, log in through the browser, and describe the task in plain language. The first thing I do in a new project is the /init command: the agent writes a CLAUDE.md file describing the structure, conventions, and test commands, and from then on reads it at the start of every session. It’s the single file that most improves work quality.
Where It Works: Terminal, Editor, Browser
The base is the terminal, but the same tool has extensions for VS Code and JetBrains (changes appear as diffs in the editor), a browser version launched from claude.ai, and a desktop app. It connects to external tools (database, browser, Jira, custom APIs) through MCP, an open standard I described in the article about building agents with Claude and MCP. It splits larger tasks among sub-agents and saves repeatable actions as custom slash commands.
Pricing Details
Claude Code is included in the subscription price: Pro is $20 per month, Max starts at $100. Limits are counted in five-hour windows and shared with chat, so an intense coding day on Pro ends with a limit message after a few hours; Max gives five or twenty times more. With API billing, you pay for tokens according to model rates, and an hour of agent work on a large repository typically costs a few dollars. All plans and API rates are detailed in the article how much Claude costs.
Claude Code, Cursor or Codex
| Claude Code | Cursor | Codex (OpenAI) | |
|---|---|---|---|
| Form | agent in terminal and editor | code editor with agent | agent in terminal, cloud and ChatGPT |
| Who controls | agent executes task, you approve | you write, agent helps and executes | agent executes, also in background on server |
| Entry price | in Pro plan ($20) | Hobby free, Pro $20 | in ChatGPT Plus and Pro plans |
| Best for | large changes in existing project, refactors, tests | daily code writing with preview | tasks assigned in background, several at once |
In practice, many developers use two at once: Cursor as an editor, Claude Code for tasks that should “do themselves.” A comparison on the same task is in the article Cursor versus Claude Code.
Five Habits That Save Money and Nerves
I start with planning mode and read the plan before the agent touches files. I divide tasks into ones that fit in a single session. I make it run tests after every change and don’t approve when they’re red. I work on a separate Git branch so rolling back is one command. And I set a cost threshold in settings when paying through API.
Reader Questions
Is Claude Code free?
No. It requires a Pro or Max plan or an API key with funded balance. Only the installation is free.
Does Claude Code work in other languages?
Yes, commands and conversation can be in any language, and it will write code comments in the language you request. Variable names and error messages are best left in English, as that’s what the rest of the ecosystem uses.
Do I need to know how to code?
For simple scripts and sites, no; for working on a real project, yes, at least at the level of reading code and understanding what the agent proposes. Without that, you can’t distinguish a good change from one that will pass tests and break production.
When to Choose Something Else
If you prefer working visually in an editor rather than in the terminal, Cursor will be more natural. If you can’t code and want a finished app rather than code to review, tools like Lovable are appropriate. And if you’re looking for an environment without installing anything, check out the guide to Replit. I describe the model itself and its other applications in the guide to Claude.
See also: Codex or Claude Code - comparison of code agents, and MCP.
See also: Warp: terminal with AI agent · Factory.ai and Droid
More on this from QUANTSLANT
guides·Ollama: How to Run AI Models Locally on Your ComputerOllama runs AI language models on your computer for free, without an account. Installation guide for Windows and Mac, essential commands, and test results.