Claw field notebook
● last updated 2026-05-15 edit on GitHub colophon
Anthropic / Claude Code / CC.1 · 4 min read

Claude Code, plainly

Five-minute orientation to Claude Code — what it is, where it runs, what it costs, and where it shines vs Codex CLI / Aider / Cursor / GitHub Copilot CLI.

The thirty-second version#

Claude Code is a terminal coding agent. You install a CLI on your machine, run claude inside any project directory, and Claude reads your code, edits files, runs commands, and stays in the conversation until you’re done. It speaks the Model Context Protocol natively (the standard Anthropic kicked off in late 2024), so you can wire in external tools — your filesystem, a database, a browser, a Jira queue — and Claude can drive them.

Same engine ships in four other surfaces: a VS Code / Cursor / JetBrains extension, a desktop app (macOS · Windows), and a web surface at claude.ai/code. Your CLAUDE.md project files and settings carry across all five surfaces — pick whichever fits the moment.

A real example#

You’re staring at a Next.js project that’s failing to build. You open a terminal, cd into the project, type:

claude

Then ask in plain English: “the build is failing — figure out why.”

What happens:

  1. Claude reads the error from npm run build.
  2. It opens next.config.mjs, sees an experimental flag that’s been removed.
  3. It checks package.json to see what version of Next.js you’re on (15.3 — the flag landed in 14.2).
  4. It edits the config file, removes the flag, runs npm run build again.
  5. Build passes. Claude tells you what it did.

No screen-sharing with a colleague, no Stack Overflow rabbit hole. The agent ran in your terminal, with full read-write access to your repo, scoped to that one session.

Why this matters#

The interesting bit isn’t “AI writes code” (every editor has that now). The interesting bit is the loop tightens: the agent reads → decides → runs → reads-the-result → decides-again, without you copying snippets back and forth. That’s why people who use Claude Code daily talk about “delegating an hour” of work, not “getting autocomplete.”

Three things that fall out of this design:

  • It’s your repo, on your machine. Files are edited locally. Commits happen through your normal git workflow. The agent doesn’t push, doesn’t open PRs unless you ask. What does cross the wire: the model call to Anthropic’s API includes whatever code, file contents, and conversation context Claude needs to answer — that’s how any LLM-backed tool works. The point isn’t “your code never leaves your machine” (it sometimes will, as context for the model); the point is “your code lives in your local git workflow, not as state on someone else’s product.”
  • Tools are pluggable via MCP. The protocol exists exactly so you can teach Claude how to read your bug tracker, query your DB, or hit a private API — without Anthropic having to ship a specific integration for every tool that exists.
  • You can run multiple agents at once. The desktop app supports parallel sessions. The CLI supports -p (one-shot, non-interactive) mode that scripts can call. Side-by-side investigation of two branches is a real workflow.

How it compares#

Claude CodeCodex CLIGitHub Copilot CLICursor / Windsurf
ShapeTerminal CLI + IDE ext + desktop + webTerminal CLI + IDE extTerminal CLI + IDE extIDE-first (forked VS Code)
VendorAnthropicOpenAIGitHub / MicrosoftIndependent (Anysphere / Codeium)
ModelClaude familyOpenAI GPT familyMulti-vendor (Claude · GPT · Gemini · others)Multi-vendor
MCPNativeSupportedSupportedSupported (Cursor)
PricingSubscription (Pro / Max / Team) + API tierSubscription (Plus / Pro) + API tierSubscription (Individual / Business)Subscription (Pro / Business)
Best atLong agentic tasks, MCP-heavy workflowsTight OpenAI-stack workflowsMulti-model flexibility · tight GitHub integrationIn-editor flow with codebase indexing
TradeoffOne model family · pricing scales hard above ProOne model family · narrower MCP catalogueLess aggressive on multi-step agentic workIDE lock-in

Sush’s honest take (sourced, not yet tried): the multi-model surface in GitHub Copilot CLI is genuinely useful — Sush uses Claude through it daily, and switches to GPT when Claude struggles or to Gemini when long context matters. Standalone Claude Code probably wins on agentic depth + the MCP catalogue is bigger; but the multi-model flexibility is a real reason to keep Copilot CLI in the toolbox.

What it costs#

There are two ways to pay:

  1. Claude.com subscription — Pro ($17/mo billed annually [$200 up front], $20/mo billed monthly), Max ($100/mo for 5× Pro or $200/mo for 20× Pro), Team (per-seat). Higher tiers raise the per-week message + context budget. Subscription tiers are the most predictable cost if you use it daily.
  2. Pay-as-you-go via Anthropic Console — set a credit balance, get charged per-token. Useful for occasional heavy bursts; expensive if you forget to monitor.

You can also point Claude Code at a third-party provider — Amazon Bedrock, Google Vertex AI, or a self-hosted Claude proxy — via env vars. Useful for enterprise/governance reasons; check /docs/en/third-party-integrations for the current list. Pricing on Bedrock/Vertex is the cloud’s own per-token rate, not Anthropic’s direct rate.

⚠️ Pricing changes faster than this page. Always check anthropic.com/pricing before committing.

What’s “load-bearing” in your project#

Once you start using Claude Code seriously, three files matter:

  • CLAUDE.md — project-level instructions. Style guide, architecture notes, “always run npm run lint before committing,” “the test command is pnpm test:unit.” Claude reads this on every session.
  • ~/.claude/settings.json — your personal config (model preferences, default thinking budget, MCP server allowlist).
  • .claude/ directory in your project — slash command shortcuts, subagent definitions, project-scoped MCP server pins.

These are the files you’ll edit most as you tune the experience. See §CC.4 Auth + API key for the credentials side and §CC.5 MCP integration for plugging tools in.

What to do next#

Sources