Why VS Code + Claude Code Became Our Default Over Cursor — The Setup Anthropic’s Own Engineers Use

TL;DR: Anthropic’s engineering team writes >80% of merged code via Claude Code in VS Code (not Cursor). We mirrored their stack: VS Code for editing/debugging, Claude Code terminal agent for multi-file refactors, architecture, autonomous loops. Cursor Composer context loss + $200 Ultra tier made the switch obvious. Full setup guide →

The Context

Two-dev team, 120 hrs/mo AI coding. Jan–May 2026: Cursor Pro ($20) primary, Anthropic API direct ($150–300/mo) for agentic work. Frustrations: Composer loses context across files >500 lines; no terminal autonomy (can’t run tests, migrate DB, deploy); $200 Ultra tier for “more agentic” felt like pricing hostage. May 2026: Anthropic revealed Claude Code writes 80%+ of their production code.

What We Tested

Tool	Use Case	Verdict	Why
Cursor Pro	Multi-file refactors, Composer agent	❌	Context loss >500 lines; no terminal autonomy; $200 Ultra for basics
VS Code + Claude Code	Editing + terminal agent (refactors, tests, infra)	✅	Native terminal; persistent context; cloud Routines; sub-agents
VS Code + Copilot Chat	Inline completions, quick questions	❌	Token pricing broke economics; no agentic loops
Windsurf	Dedicated editor, agentic flows	⚠️	Good agentic UX; but locked into their editor; $15/mo adds up

The Pivot Point

May 28, 2026: Debugging a payment webhook race condition. Cursor Composer: 3 context resets in 45 min, couldn’t run npm test to verify fixes. Switched to VS Code + cc (Claude Code alias). Single session: wrote test, ran it, fixed race condition, ran migration, deployed to staging — zero context loss, terminal autonomy throughout. Realized: we were fighting Cursor’s architecture; Claude Code is the terminal.

What We Use Now

Daily Driver: VS Code (stable) + Claude Code (terminal agent)

# ~/.zshrc aliases
alias cc='claude-code --model sonnet'                    # Default: Sonnet in credit pool
alias cc-opus='claude-code --model opus --reason'        # Opus: requires reason flag
alias cx='codex'                                         # Codex for research/scaffolding
alias cursor='cursor'                                    # Cursor Pro: <30 min surgical TS edits

VS Code Extensions (minimal, purposeful):

• anthropic.claude-code — official, shows credit usage in status bar
• github.copilot — disabled (kept for emergencies only)
• ms-vscode.typescript-next — TS nightly for better IntelliSense
• esbenp.prettier-vscode — format on save

Workflow Protocol:

1. Architecture / multi-file / infra → cc in terminal (autonomous, runs tests, edits files)
2. Research / scaffolding / docs → cx (Codex persistent agent)
3. Quick TS/React edits (<30 min) → cursor (Composer surgical)
4. Debugging / interactive → VS Code native debugger + cc for fix proposals

Cloud Routines (Claude Code): Scheduled cc runs for dependency updates, security audits, test generation — configured via claude-code routine add --cron "0 2 * * 1" --prompt "Audit deps, generate tests, open PR"

When You’d Choose Differently

• Team refuses terminal workflow: Cursor/Windsurf dedicated editors lower friction
• Heavy Opus needs (>50%): Credit pool economics break; direct API or o1-pro may win
• Windows without WSL2: Claude Code terminal UX degrades; Cursor/Windsurf better native
• Solo dev, <20 hrs/mo: Cursor Pro $20 simpler than credit pool $100 minimum

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	5	Best agentic UX 2026; terminal-native = workflow-native
Ease of Use	3	Terminal learning curve; alias discipline required
Value	5	$100/mo fixed vs $200–800 variable; 80% Anthropic dogfood proof
Support	4	Discord active; claude-code usage CLI excellent; Routines beta

---

This is part of our AI Coding Tool Evaluation series. See full setup: VS Code + Claude Code 2026: The Anthropic-Engineer Stack