Claude Code Review

Claude Code is Anthropic's official AI pair-coder. It runs in your terminal — not in a separate IDE — and treats the CLI as the first-class surface for AI-assisted development. You give it a goal, it plans the work, edits multiple files, runs commands, observes the output, and iterates.

It's the closest thing to having a senior engineer in your terminal who reads the codebase, asks clarifying questions, and ships. The plan mode is the differentiator: instead of dropping a code suggestion and hoping you accept it, Claude Code shows you the multi-step plan first, lets you steer, then executes.

How it differs from Cursor and Copilot

Cursor is an IDE. It's VS Code plus AI inline. You see your editor and the AI augments it. Claude Code skips the IDE — it's a CLI that reads, edits, and runs. If you're a developer who lives in tmux + vim + git, Claude Code feels native. If you live in VS Code, Cursor probably wins ergonomically.

Copilot is autocomplete. It predicts the next token. Claude Code is agentic — it decomposes goals, runs tools, and operates over multiple files. Different mental models, different price points.

Pricing structure (last tested May 2026)

There are five tiers. Free is for trial only — rate limits make any real work impractical. Pro at $20/mo is the entry point and the right pick for most indie devs. Max 5× ($100/mo) is what you want if you're using Claude Code for several hours daily. Max 20× ($200/mo) is for operators running it as their primary IDE-replacement. Team is per-seat with shared caps and central admin.

The pricing tracks model usage rather than seats, which is unusual for developer tools and worth understanding before you scale up. If you hit Pro limits midway through a refactor, you wait for the rate-limit reset or upgrade — there's no usage-based pay-as-you-go option.

What works

• Plan mode actually plans. It shows you the multi-step approach before executing, lets you reject or revise, and doesn't bury you in 17 file diffs you have to scan.
• Sub-agents for parallelism. You can spawn worker agents to handle independent sub-tasks (run tests, scan a directory, draft a migration) while the main agent keeps planning. Real time-savings on large refactors.
• Model choice. Opus for the hardest reasoning, Sonnet for fast iteration, Haiku for cheap broad tasks. You pick per task or let the tool route automatically.
• MCP support is native. Connect any MCP server (database, GitHub, Linear, your custom one) and Claude Code uses it as a tool. This is where it pulls ahead of Cursor for ops-heavy work.
• Git-aware. It reads your branch state, can diff against a target, drafts commit messages from the staged changes. Small details, big day-to-day savings.
• Terminal-first respects your existing workflow. tmux, vim, fish — whatever you already use, Claude Code slots in.

What doesn't

• No visual canvas. If you're building UI and you want to see what you're editing, you'll still want a real IDE open. Claude Code is great at writing components — less great at iterating on their visual output.
• Pricing scales fast. $20/mo Pro is fair, but if you actually use it daily, you'll hit Pro limits and the next tier is $100. The gap is real.
• Claude-only. No GPT, no Gemini. Anthropic's models are excellent — but if you've benchmarked GPT-5 or Gemini Pro for a specific task and it wins, you can't route to them inside Claude Code.
• Onboarding curve. The first day with plan mode + sub-agents feels like over-engineering. The third day you realize it's why you finish refactors that used to die at file 4.

I've used Claude Code as the primary AI pair-coder on every Vibetoolstack-related codebase since February 2025: this Astro/Sanity site, the Cropsharing MVP, several smaller side-builds, and the internal tooling I run for content production. Twelve months, three production codebases, hundreds of sessions.

The kind of tasks where it consistently wins

Multi-file refactors. The classic example: adding a new field to a Sanity schema, then propagating that field through the GROQ query, the Astro template, the TypeScript type, and three component files. With Cursor I'd do this in 4-5 separate Composer prompts. With Claude Code I describe the goal once, it plans the cascade, and I review one diff.

MCP-driven workflows. I run a custom MCP server for VTS internal tooling — keyword research, content audits, deployment hooks. Claude Code calls into it the same way it calls into the file system. That's a categorically different developer experience than "copy the data into the chat".

Sub-agents for content audits. Spawning 4 agents in parallel to scan different content directories for issues, while the main agent compiles the report. The first time this worked end-to-end without hand-holding, I bumped the VTS score by 5.

Where it stumbles

Visual UI iteration. When I'm tweaking padding, font-weight, or a layout's responsive breakpoints, I'm faster in DevTools + Cursor inline-completion than describing the change to Claude Code. It's a tooling-fit problem, not a capability problem — Claude Code can absolutely write the CSS, but the loop of "change → see → adjust" is slower in CLI than in a visual editor.

Long-context drift on hour-3 sessions. Once a session hits 600K-700K tokens of context, the model starts repeating itself or losing earlier constraints. The fix is to wrap a discrete task and exit the session, not to push the context. Anthropic shipped Opus 4.7 with 1M context; this got 80% better but isn't fully solved.

Real numbers from 12 months of use

Pro at $20/mo covered me for the first 4 months while I was prototyping. Once I started shipping daily, I hit the rate limits enough times that I moved to Max 5× ($100/mo) and stayed there. Max 20× was tempting but I haven't needed it — the bottleneck for me is the work I queue, not the tool's rate limits.

Output: this Vibetoolstack site (Astro + Sanity, ~80 components), Cropsharing MVP (full-stack, Supabase + Stripe + auth), and dozens of internal scripts. About 75% of the code shipped in those projects passed through Claude Code at least once.

When Claude Code is the right pick

Indie founders shipping AI-native products. You don't have a co-founder. You need to ship features, not draft them. The plan mode + sub-agents combination is most valuable here — you compress what would take a week of manual coding into days.

Developers maintaining medium-to-large codebases. The bigger your codebase, the more multi-file refactors hurt. Claude Code's context-aware editing and plan mode pay off most when there are 5+ files involved per change.

Operators running custom MCP servers. If you're already invested in the MCP ecosystem (database tools, ops automations, internal APIs), Claude Code is the IDE that natively speaks MCP. No plugin gap.

CLI-fluent developers. If your daily setup is tmux + vim + fzf + git, Claude Code drops in without changing your workflow. You don't open a separate window — it lives in the same terminal.

When something else wins

Visual UI work. v0 for shadcn-flavored React components, Cursor inline-completion for live tweaks, or just plain DevTools.

Microsoft-stack teams. If your team is on VS Code with Copilot business licenses and Azure infra, the integration cost of Claude Code is real and Copilot is good enough at autocomplete.

Free-tier prototyping. The Free tier rate limits make any real session impractical. If you're cost-sensitive and just exploring, Cursor's free tier or even regular Claude.ai web is friendlier.

Within the AI coding category, the live alternatives all do something different — Claude Code isn't strictly better than each, just optimized for a different surface.

• Cursor — IDE-native, the strongest direct competitor. Pick Cursor if you live in VS Code and want AI inline. Pick Claude Code if you live in the CLI.
• Windsurf — Codeium's IDE with the Cascade agent. Closer to Cursor than to Claude Code. Worth watching for the agent layer if you don't want to leave the IDE.
• GitHub Copilot — the OG. Strong autocomplete, weaker on agentic multi-file work. Best if you're locked into Microsoft's ecosystem.
• Lovable / Bolt / v0 — different category (prompt-to-app). Pick one of these if you want to ship a working app from a prompt, not edit a codebase.

Full alternatives breakdown with one-liners and pricing on the dedicated alternatives page.

Is Claude Code worth $20/mo?

If you write code daily and you've been using Cursor or Copilot, yes — even just for plan mode and sub-agents on multi-file work. If you're an occasional developer who codes 1-2 hours a week, the Free tier or Cursor's free tier covers you.

How does Claude Code differ from claude.ai (the chat web app)?

claude.ai is a chat interface for general conversation. Claude Code is a CLI agent with file-system access, command execution, plan mode, sub-agents, MCP support, and git awareness. Same model, very different surface.

Can Claude Code run offline?

No. It calls Anthropic's API for every model invocation. There's no local model option. If you need offline, look at locally-hosted Ollama + a CLI wrapper, but expect significantly weaker output.

Does it support GPT or Gemini models?

No. Claude Code only routes to Anthropic models (Opus, Sonnet, Haiku, plus the Claude Code Specialist). If you need cross-model routing, Aider supports it; Cursor lets you bring your own model keys.

Is the codebase data sent to Anthropic?

Files Claude Code reads or edits are sent to the API as context. Anthropic's API does not train on customer data. SOC 2 Type 2 certified. For sensitive codebases, run Claude Code in an environment where you control what files it can access (the CLI respects .gitignore by default).

How much does it actually cost in real use?

Pro at $20/mo for 4-8 hours a week of active use is fine — you'll occasionally hit limits but they reset every 5 hours. For daily developers (4+ hours/day), expect to land on Max 5× at $100/mo. Max 20× at $200 is for operators running it as their primary work surface.