Claude Code is Anthropic's autonomous coding agent. It runs in your terminal, reads your entire codebase, writes code, executes shell commands, and handles multi-step tasks without you supervising each step. It's not autocomplete. It's not a chatbot with code blocks. It's an agent that does the work.
Everything in this guide comes from daily hands-on use at Northbase and across other businesses, not documentation rehashed.
What Claude Code Actually Is
Claude Code is a command-line tool you install via npm. You run it in a terminal. It reads your project files, understands the codebase, and then you tell it what to do in plain language. It plans the work, edits files, runs commands, checks its own output, and iterates until the task is done.
The key thing that separates it from every other AI coding tool: it doesn't need a GUI. No editor. No browser window. No tabs. It runs in a terminal, which means it runs over SSH, inside tmux sessions, on headless servers, through Telegram bots, on a Mac Mini in your cupboard. You can text it an instruction from your phone and check the results later.
That's not a minor technical detail. It's the entire reason Claude Code is different from tools like Cursor, GitHub Copilot, or Windsurf. Those tools need you sitting at a computer with an editor open. Claude Code doesn't.
How Claude Code Works
You install it globally:
npm install -g @anthropic-ai/claude-code
You navigate to a project directory and run claude. It reads your project structure. Then you talk to it.
A typical interaction isn't "write me a function." It's more like: "Add a contact form to this Astro site that stores submissions in Cloudflare D1 and sends me a Telegram notification. Include server-side validation and a success state." Claude Code will read your existing code, understand the project structure, create the necessary files, write the worker function, update the form component, and test the build. Multiple files. Multiple steps. One instruction.
It works because Claude Code has access to your entire local environment:
- File system - reads and writes any file in your project
- Terminal - runs shell commands, build tools, tests, git
- Context - reads CLAUDE.md files for project-specific instructions
- Network - full network access for API calls, deployments, package installs
The CLAUDE.md file is worth mentioning specifically. You put a markdown file in your project root that tells Claude Code how your project works - the tech stack, conventions, gotchas, what not to touch. It reads this every session. It's like giving a new developer onboarding notes, except this developer actually reads them.
Recent additions worth noting: Claude Code now has automatic memory - it records and recalls context as it works across sessions. There's a research preview of agent teams for multi-agent collaboration. And it integrates directly into VS Code with session management and plan review, though the terminal remains the primary interface for autonomous work.
Which Model to Use
Claude Code runs on Anthropic's Claude models. The two that matter:
Opus 4.6 is the model you want for serious work. Complex reasoning, architectural decisions, multi-file refactors, anything where the agent needs to hold a lot of context and make judgment calls. It's slower and more expensive. It's worth it. When I advise clients on which model to use as an agent orchestrator, the answer is always the smartest one available. Right now that's Opus.
Sonnet 4.6 is faster and cheaper. Good for simpler tasks - quick edits, straightforward features, boilerplate. If you know exactly what you want and the task doesn't require much reasoning, Sonnet saves you money. But don't use it for complex work. The quality difference is real.
| Opus 4.6 | Sonnet 4.6 | |
|---|---|---|
| Best for | Complex reasoning, architecture, multi-file | Simple edits, boilerplate, quick tasks |
| Speed | Slower | Faster |
| Cost per task | $10-50 | $2-10 |
| Context handling | Excellent across long sessions | Good for shorter sessions |
| Judgment quality | Highest available | Good, not great for edge cases |
My take after months of daily use
Even with Codex 5.4 now available, Opus 4.6 is still the best main agent brain for your business. I've been running it daily for as long as it's been available. Opus is way better at front-end design than Codex. But the real difference is behaviour - its own internal personality is way more moldable than ChatGPT's Codex. When I specify how I want it to work, it actually listens. I get better coding results with less interaction. The depth of reasoning, the ability to hold complex architectural context across long sessions, the judgment calls on ambiguous problems - nothing else matches it right now. Sonnet 4.6 is faster but not as thorough. Sonnet is only worth using if you don't want to pay for the Max plan at $200 USD a month (roughly $300 AUD).
And if you're seeing people online recommending Qwen 3.5 as a main agent brain - don't. I tried it. Extremely slow, not smart, reminded me of ChatGPT 3.5 with its overuse of emojis. The real problem: my expectations were misaligned after sitting on Opus 4.6 constantly, then going back to Qwen expecting similar quality. There's a huge gap between what an open-source model benchmarks on a leaderboard and how it actually performs when you need it to reason across a real codebase, make real architectural decisions, and recover from real errors. Qwen 3.5 couldn't do any of that competently. The people recommending it either haven't used it for serious agentic work or have very different standards to mine.
The model you use as your agent's brain is the single most important decision in the whole stack. Cheap out there and everything downstream suffers. This is not the place to save money.
One more thing: the most overhyped trend in AI right now is using a different AI app for everything. You can use Claude Code with Opus 4.6 as an agent operating system and it does everything those $50/month apps offer. It's very easy to consolidate right now. The personalisation factor with Claude Code is huge - it knows your entire business, your clients, your tools, your knowledge, your tone of voice. Stock standard ChatGPT on your phone is a completely different thing to Claude Code running actively on its own computer with full business context.
What Claude Code Is Good At
Full-stack development. Tell it to build a feature end-to-end. It writes the frontend component, the API endpoint, the database migration, the types. It understands how the pieces connect because it reads your whole codebase first.
Refactoring. "Rename this service and update every file that imports it" is a five-second instruction that handles what would be 20 minutes of find-and-replace. But the real value is larger refactors - restructuring a module, changing an architectural pattern, migrating from one approach to another.
Debugging. Paste an error. It reads the relevant files, traces the problem, and fixes it. Often it'll find the root cause was three files away from where the error surfaced. That kind of cross-file reasoning is where Claude Code shines.
Multi-step workflows. "Build the page, run the build, take a screenshot, check if it looks right, fix the responsive issues, rebuild." That's one instruction. Claude Code chains the steps together, checks its own work, and iterates. This is agentic AI in practice.
Codebase understanding. "How does authentication work in this project?" Claude Code reads the relevant files and gives you an accurate explanation. Useful for onboarding into unfamiliar codebases or understanding code you wrote six months ago and forgot about.
What Claude Code Is Not Good At
I'm not going to pretend it's perfect. Honesty matters more than hype.
Visual design. It can write CSS. It can't see the result. It doesn't know if a page looks good. You need to review visual output yourself. Tools that can take screenshots and feed them back help, but Claude Code's native workflow is text-based.
Tasks that need real-time human judgment. If the task requires "use your taste" or "make it feel right," you're going to be going back and forth. Claude Code is best when the success criteria are clear. Ambiguous creative direction leads to iteration loops.
Extremely large codebases without guidance. On a massive monorepo with hundreds of thousands of lines, Claude Code needs your help navigating. The CLAUDE.md file and clear instructions about which parts of the codebase to focus on make a big difference. Without guidance, it can spend a lot of tokens reading files it doesn't need.
Tasks where you need to watch every change. If you want to see diffs appear in real time and approve each one visually, Cursor is better for that workflow. Claude Code shows you what it did after it's done. That's the trade-off of autonomous execution. Cursor's new Automations feature also now offers cloud-based unattended execution, so the headless advantage is narrowing.
Claude Code Pricing
Claude Code itself is free to install. You pay for the API usage.
Pricing is based on tokens - how much text the model reads (input) and writes (output). In practice:
- Simple task (quick edit, small fix): $2-5
- Medium task (new feature, multi-file changes): $10-25
- Complex task (architectural refactor, full-page build): $20-50
- Heavy day (continuous autonomous work): $50-100+
Claude Code Max is Anthropic's subscription option. $100/month gives you 5x the usage of the Pro plan. $200/month gives you 20x. Both include access to Claude Code and Cowork (multi-step task delegation). Usage limits reset weekly. If you're using Claude Code daily, Max gives you predictable costs instead of surprise API bills.
The way to think about cost: compare it to your time, not to other software subscriptions. If a $20 API bill replaces three hours of manual work, that's not expensive. If you're using it for five-second edits you could do faster by hand, that's waste.
Claude Code vs the Competition in 2026
AI coding tools have converged hard. Every major player now offers some form of autonomous agent. Here's where things actually stand as of March 2026:
| Claude Code | Cursor | OpenAI Codex | GitHub Copilot | |
|---|---|---|---|---|
| Type | Terminal agent | IDE + cloud agents | Terminal + cloud agent | IDE extension |
| Autonomous | Yes, fully headless | Yes, IDE + cloud VMs | Yes, cloud sandboxes | Limited |
| Runs unattended | Yes | Yes (Automations) | Yes (cloud) | No |
| Multi-file | Yes | Yes | Yes | Limited |
| Parallel tasks | One at a time | Multiple (Automations) | Multiple cloud agents | No |
| Model | Claude only | Multi-model | GPT-5.4 | Multi-model |
| Pricing | API or Max $100-200/mo | $20-40/mo | $20-200/mo | $10-19/mo |
What's changed recently: Cursor launched Automations in March 2026 - always-on cloud agents that trigger from Slack, GitHub, Linear, and webhooks. Each runs in its own VM and produces merge-ready PRs. This closes the gap with Codex on parallel/unattended work. OpenAI shipped GPT-5.4 with native computer-use capabilities and 1M token context. Claude Code added agent teams (research preview) for multi-agent collaboration and automatic memory across sessions.
For the detailed breakdowns: Claude Code vs Cursor and Claude Code vs ChatGPT/Codex.
The short version: Claude Code's advantage is full local system access and the deepest integration with Claude's reasoning. Cursor now has both IDE-based and cloud-based agent execution. Codex has parallel cloud execution powered by GPT-5.4. None of them is universally better. It depends on how you work and which model's reasoning you trust most.
How I Actually Use Claude Code
I run Claude Code on a Mac Mini that stays on all day. I have a Telegram bot that pipes messages to Claude Code sessions. I text instructions from my phone. The agent does the work. I review results when I'm ready.
A typical morning: I text "audit the SEO on the advisory page and fix anything you find." Claude Code loads the SEO expert instructions from a markdown file, runs the audit, identifies issues, fixes them, rebuilds the site, and sends me back a summary. I didn't open a terminal. I didn't open a browser. I reviewed the summary on my phone and approved the deploy.
This is what I mean by a business agent operating system. Claude Code is the execution engine. The structure around it - specialist instruction files, client workspaces, a Telegram bot, persistent memory via markdown files - turns a single tool into a system. The system is deliberately simple. Folders, files, and a bot. No framework.
That's the real unlock with Claude Code. It's not about writing code faster. It's about removing yourself from the execution loop entirely and focusing on judgment and direction instead. The person who runs this kind of system is an agent operator - and it's a role every business running AI will need.
Making AI tool decisions for your business?
AI advisory for founders and business owners who need direct answers on AI tools and implementation. From someone who builds with these tools every day.
AI advisory services →Getting Started with Claude Code
If you want to try Claude Code, here's the honest path:
1. Get an Anthropic API account. Go to console.anthropic.com, create an account, add credits. Start with $20. That's enough for a solid day of experimentation.
2. Install Claude Code. npm install -g @anthropic-ai/claude-code. You need Node.js installed.
3. Start with a real project. Don't test it on toy examples. Open a real codebase you're working on. Navigate to the directory. Run claude.
4. Start with clear, concrete tasks. "Add a 404 page to this project" is better than "make the site better." Claude Code excels when the success criteria are unambiguous. As you get comfortable, you can give it broader instructions.
5. Create a CLAUDE.md file. Tell it about your project. The tech stack, the conventions, what matters. This is the single best thing you can do to improve Claude Code's output quality.
6. Use Opus for anything complex. Don't try to save money by using a weaker model for hard tasks. The cost difference between Opus and Sonnet is small compared to the time you waste fixing bad output from a model that wasn't smart enough for the job.
What People Get Wrong About Claude Code
"It's just for developers." Claude Code writes code, but the person using it doesn't have to be a developer. I've seen business owners use it to build landing pages, set up automations, and manage their web presence. This is why WordPress is pointless for new websites - Claude Code can build a faster, cheaper static site in an afternoon. You need to be comfortable in a terminal, but you don't need to be a programmer. What you need is the ability to describe what you want clearly. That's a communication skill, not a coding skill.
"It replaces developers." It doesn't replace developers. It replaces a lot of the work developers do. The judgment, architecture, and taste still come from a human. Claude Code handles the execution. If anything, it makes good developers far more productive and makes the gap between good and mediocre developers wider.
"It's too expensive." Compared to what? If a $30 API bill replaces half a day of manual work, that's the cheapest employee you'll ever hire. The people who think it's expensive are comparing it to free tools, not to the time it saves.
"I should wait for it to get better." It's already good enough to be useful. Waiting means your competitors learn it first. Every month you're not using it is a month someone else is building that muscle memory and you're not.