Why We Run VS Code + Claude Code Instead of Cursor for Greenfield Work — The Setup Anthropic Uses Internally

TL;DR: Anthropic writes >80% of merged code with VS Code + Claude Code (not Cursor). We mirrored their terminal-native config — cc --allowedTools for autonomous loops, Routines for reusable workflows — and reduced greenfield feature time 35%. Full setup guide →

The Context

Two-dev team building client projects (Next.js/Supabase/Stripe). Previously: Cursor as daily driver for everything. June 2026: Anthropic revealed Claude Code writes 80%+ of their production code. Key difference: terminal-native autonomy — Claude Code runs in your shell, keeps dev servers alive, executes verification loops (write → test → fix → re-run) without leaving the terminal.

What We Tested

Configuration	Use Case	Verdict	Why
Cursor (default)	Greenfield API endpoint	❌	Composer loses terminal state; no autonomous test-fix loop
VS Code + Claude Code	Greenfield feature (payments webhook)	✅	cc --allowedTools "Bash,Edit,Write" ran 45-min unattended test-fix loop
VS Code + Claude Code	TypeScript-heavy refactor (14 files)	⚠️	No LSP integration; Cursor catches type errors mid-edit
Cursor + Copilot	Inline completions + refactors	✅	Best of both for type-heavy work; $40/mo combined

The Pivot Point

June 8, 2026: Built new Stripe webhook endpoint (greenfield). Tried in Cursor — Composer generated code but couldn’t run stripe listen tunnel + test loop autonomously. Switched to Claude Code: cc --allowedTools "Bash,Edit,Write,Task" "Build webhook endpoint with idempotency, write tests, run until green". 45 minutes later: 247 tests passing, 3 files patched, $0.12 API cost. Cursor would have required manual test runs each failure.

What We Use Now

VS Code + Claude Code Config (.vscode/settings.json + ~/.claude/settings.json):

// ~/.claude/settings.json
{
  "allowedTools": ["Bash", "Edit", "Write", "Read", "Task", "Glob", "Grep"],
  "defaultModel": "sonnet",
  "routinesPath": ".claude/routines"
}

Team Protocol:

• Greenfield features → cc (Claude Code) with Routine library (.claude/routines/): fix-tests.yaml, scaffold-api.yaml, migrate-db.yaml
• TypeScript refactors → Cursor (LSP catches errors mid-edit)
• Inline completions → Copilot (always on)

Routine Example (.claude/routines/fix-tests.yaml):

name: fix-tests
description: "Run tests → on failure, analyze → patch → re-run until green"
steps:
  - run: "pnpm test"
  - on_failure: "analyze AND patch AND re-run until passing"
  - max_iterations: 5

When You’d Choose Differently

• TypeScript-first codebases: Cursor’s LSP integration catches type errors while you edit — Claude Code only sees errors after running tests.
• Teams needing IDE-integrated chat: Windsurf ($15/mo) keeps Cascade agents inside a VS Code fork.
• Pure terminal/Tmux workflows: Aider (BYOK, git-native) or Codex persistent agent if you prefer chat UI.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	4.5	Best terminal-native autonomy; changes how you think about AI coding
Ease of Use	3	CLI-first; requires Routine authoring to unlock full value
Value	5	$100/mo credit pool covers heavy Sonnet usage; 40–60% savings vs API
Support	3	Discord only; Routines documentation sparse

---

This is part of our AI Coding Tool Evaluation series. See full setup guide: VS Code + Claude Code: The Anthropic Internal Setup