Claude Code Developer Guide

Claude Code is Anthropic’s official command-line interface for Claude. It’s a Node.js CLI you install with npm i -g @anthropic-ai/claude-code and run from your terminal inside any repo. Unlike the chat-window claude.ai, Claude Code reads your real source tree, edits files in place, runs your tests, executes shell commands you authorise, and ships PRs end-to-end. It’s designed to be an AI pair-programmer — you steer, Claude does the keystrokes. Built and maintained by Anthropic, the same company that makes Claude itself.

Three key differences:

• Vs. Copilot: Copilot is autocomplete-in-IDE. Claude Code is a CLI agent that plans, edits multiple files, runs tests, and commits PRs — not just inline suggestions.
• Vs. Cursor: Cursor is an IDE (VS Code fork) with chat. Claude Code is editor-agnostic — works in your existing VS Code, JetBrains, Vim, or terminal. Same agent, no IDE lock-in.
• Vs. Aider: Aider is similar in shape (CLI pair-programmer) but provider-neutral. Claude Code is Anthropic-native, so you get first-party access to Claude features (extended thinking, prompt caching, computer use, MCP) without a translation layer. Sub-agents and hooks are also first-party.

For Magento & Hyvä work specifically, Claude Code wins on: real codebase reads, MCP-database access, hook-gated safety, and the slash-command library.

By default, the latest production-tier Anthropic model — currently Claude Opus 4.7 for heavy coding tasks and Claude Sonnet 4.6 for fast routine work (May 2026). You can pin specific models with --model or set defaults in your .claude/config.json. Extended thinking mode kicks in automatically for complex tasks (architectural decisions, debugging). You can also route Claude Code via AWS Bedrock or GCP Vertex AI if your org requires cloud-vendor billing — the CLI picks up ANTHROPIC_BEDROCK_BASE_URL / ANTHROPIC_VERTEX_PROJECT_ID env vars.

Three steps, ~60 seconds total:

1. Install the CLI: npm i -g @anthropic-ai/claude-code (requires Node.js 18+).
2. Authenticate: Run claude login — opens browser, signs in with your Anthropic account (Pro / Team / API). Or set ANTHROPIC_API_KEY env var if you’re using API direct.
3. Run it: cd into any repo, type claude, ask your first question. That’s it.

For corporate setups behind a proxy, see the 5-step setup section above. For Magento-specific install + CLAUDE.md, see our Magento playbook.

CLAUDE.md is a Markdown file at your repo root that holds your project rules — coding standards, “never edit vendor”, “always write MFTF tests”, deploy flow, naming conventions, links to internal docs. Claude Code re-reads it at the start of every session, so your team’s rules become muscle memory for the AI. Write once, AI follows forever.

A good Magento CLAUDE.md includes: tech stack (Magento 2.4.x + Hyvä), “no vendor edits”, “plugin-over-rewrite”, “db_schema not raw SQL”, “MFTF tests required”, deploy commands, and a list of internal modules with their purpose. Sub-folders can have their own CLAUDE.md files that compose with the root one — useful for monorepos.

Different shape, similar goal. Cursor @-mentions reference files / docs / web pages inline in chat. Claude Code slash-commands are reusable prompt templates stored as Markdown files in .claude/commands/. When you type /scaffold-module, Claude Code injects the contents of .claude/commands/scaffold-module.md as your prompt — with optional argument substitution.

Slash-commands are repeatable workflows — you write the prompt once (“scaffold a Magento module with these conventions, db_schema, registration, MFTF test stub”), then any team member runs it with one keystroke. Claude Code also supports arguments ($ARGUMENTS placeholder) and file references (@path/to/file) inside command files, so you get the best of both worlds.

The Model Context Protocol (MCP) is an open standard (open-sourced by Anthropic) for connecting AI assistants to external tools and data sources. An MCP server exposes a specific resource — GitHub, Linear, Sentry, your Postgres database, your Magento DB, your design system — through a standardised interface that Claude Code can call.

You wire MCP servers in .claude/mcp.json. Once wired, Claude Code can read your GitHub issues, query your Magento DB, fetch Sentry stack traces, file Linear tickets — all without you copy-pasting context into chat. Crucially, your API keys live in MCP server config, not in your prompts — so they don’t leak into your conversation logs. Anthropic + the community ship 100+ official MCP servers; building your own is a 50-line Node script.

Hooks are shell scripts that run before or after Claude Code performs an action — editing a file, running a Bash command, finishing a tool call. They’re your safety net.

• PreToolUse hooks can block dangerous actions: refuse to rm -rf /, refuse to push to main branch, refuse to write into vendor/, refuse to edit prod credentials. Hook returns non-zero exit, action is cancelled, you stay safe.
• PostToolUse hooks auto-run lint / format / tests after every file edit. phpcs + phpstan + composer validate on every Magento file change — built into the CLI, not bolted on.
• Stop hooks run when Claude finishes a session — great for auto-committing or running test suite.

Configure hooks in .claude/settings.json. They’re the difference between “cool AI demo” and “production-grade workflow.”

Sub-agents are parallel Claude Code instances spawned by your main session, each with its own context window, working independently on a task you delegate. Useful when:

• Parallel work: spawn three sub-agents to audit, build, and test in parallel — they each get a fresh context window, so a 30-min audit doesn’t blow your main session’s context.
• Long-running tasks: a sub-agent can grind on “port these 80 Luma blocks to Hyvä” for an hour while you keep planning in your main session.
• Specialised personas: a sub-agent with a stricter CLAUDE.md (e.g. “you are a security auditor; only output findings, never edit”) gives you focused output without polluting your main agent’s instructions.

Spawn with the Task tool. They report back to the main session when done — you get the summary, not the keystroke-by-keystroke noise.

Three billing options:

• Claude Pro ($20/mo): unlimited Claude Code usage with model auto-routing. Best for solo devs — effectively flat-rate AI pair-programming.
• Claude Team ($25/seat/mo, 5+ seats): Pro features + admin controls + shared workspace + SSO. Best for engineering teams.
• Anthropic API (pay-per-token): $3/M input + $15/M output for Sonnet, $15/M + $75/M for Opus. With prompt caching enabled, repeated reads of vendor/ drop ~90% of input cost. Best for pipelined / CI / sub-agent-heavy workflows.

For most Magento dev work, Claude Pro flat rate is the right answer — you don’t want to think about per-token cost while shipping. For agentic / parallel sub-agent workflows, API with prompt caching is more economical.

No — not on commercial-tier inputs. Under Anthropic’s Commercial Terms of Service, customer inputs and outputs (Claude Pro, Claude Team, Anthropic API) are not used for model training. That’s contractual, not just a setting.

For zero-data-retention setups (regulated industries, NDA-heavy work), Anthropic supports a zero-data-retention mode on the API and Bedrock / Vertex routes — inputs aren’t logged at all past the live request. Free-tier claude.ai consumer use can opt-in to training (the consumer privacy notice differs), but Claude Code does not run on free-tier accounts. Net effect: your codebase stays in your VCS — nothing leaves except the API call you made.

Six workflows that compound massively in ecommerce dev:

1. Setup-patch generation: a /setup-patch slash-command that scaffolds idempotent DataPatchInterface patches with dependency wiring — for seeding configs, products, categories, FAQs.
2. db_schema → migration: describe the column you want, Claude Code writes db_schema.xml + db_schema_whitelist.json + a setup-patch for backfill.
3. Hyvä port: a /hyva-port-block command that takes a Luma block and spits out the Hyvä equivalent with Tailwind classes + Alpine.js.
4. MFTF + PHPUnit: features ship with tests in the same PR — non-negotiable in CLAUDE.md.
5. Sub-agent audit: nightly sub-agent that audits your repo against Magento Coding Standard + EQP, files Linear tickets via MCP.
6. SEO page builder: a slash-command that scaffolds a new Tier-1 pillar page (block + layout + template + LESS + 3 setup-patches) following your established pattern. (We use this exact one to ship our SEO landing pages.)

For the deep-dive, see our Magento playbook and Hyvä playbook.