Why We Switched Our Terminal Workflows to VS Code + Claude Code — Anthropic’s Own Stack Is 80% AI-Written

TL;DR: Anthropic’s internal data shows Claude Code writes >80% of their merged code. We moved terminal-autonomous work to cc inside VS Code, keeping Cursor only for TypeScript precision edits. Result: 40% less context-switching, zero Composer context losses on long sessions. Full setup guide →

The Context

Two-dev team, 5 active codebases. Mar–May 2026: Cursor Composer for everything. Problem: 3+ hour sessions lost running server/DB context 2–3× per session (45 min recovery each). June 2026: Anthropic revealed Claude Code writes 80%+ of their codebase. Key differentiator: terminal-native, persistent sub-agents, cloud “Routines” for repeatable workflows. We tested the hybrid: VS Code for editing/LSP, Claude Code terminal pane for autonomous loops.

What We Tested

Configuration	Use Case	Verdict	Why
Cursor Composer only	3-hr auth refactor (12 files)	❌	Lost dev server + DB tunnel 3×; 45 min recovery each
VS Code + Claude Code (cc in terminal pane)	Same refactor	✅	Terminal persisted; sub-agents handled parallel file edits
VS Code + Claude Code	Greenfield feature (new API endpoint + tests)	✅	cc wrote 85% of code; Routines captured pattern for reuse
VS Code + Cursor (no Composer)	Quick TS edits, type-error fixes	✅	LSP catches errors mid-edit; no context-switch penalty

The Pivot Point

June 10, 2026: Stripe webhook migration (14 files, running stripe listen tunnel + local Supabase). Cursor Composer: lost tunnel + DB state twice → 5.5 hrs total. Same task in VS Code + cc: terminal pane kept tunnel PID, Supabase connection, full file context. Sub-agent parallelized 14-file edit. Done in 2.8 hrs. Realization: The “IDE-integrated agent” model (Composer) loses terminal state; “terminal-native agent in IDE” (Claude Code in VS Code terminal) keeps both. We made code . && cc the default terminal-autonomous launch command.

What We Use Now

VS Code + Claude Code hybrid config (.toolcrucible/vscode-claude-code.json):

{
  "terminal.integrated.defaultProfile.osx": "zsh",
  "terminal.integrated.env.osx": { "CLAUDE_CODE_AUTO_UPDATER": "false" },
  "github.copilot.editor.enableAutoCompletions": true,
  "chat.agent.enabled": false
}

Keybindings (.toolcrucible/keybindings.json):

• cmd+shift+c: Launch cc in new terminal pane (right split)
• cmd+shift+x: Launch cx (Codex) in new terminal pane
• cmd+shift+v: Focus VS Code editor (LSP active)

Routines (saved in ~/.claude-code/routines/):

• stripe-migration.yaml: Parallel file edits, tunnel watch, Supabase seed
• auth-refactor.yaml: JWT middleware, test gen, migration safety checks
• greenfield-api.yaml: OpenAPI spec → route + handler + tests + docs

Team protocol: Terminal-autonomous = cc in VS Code terminal. IDE-integrated precision = Cursor (LSP only, Composer disabled). Persistent chat-agent = cx for archaeological refactors.

When You’d Choose Differently

• Pure VS Code loyalists: Windsurf offers IDE-integrated persistent context without terminal-native agent — but no Routines, no sub-agent parallelism.
• Teams >5 needing shared IDE config: Windsurf’s shared .windsurf config beats per-user VS Code setup.
• Strict local-only / air-gapped: Claude Code requires cloud for Routines/sub-agents; Aider + Ollama is the local terminal-autonomous alternative.
• Heavy TypeScript where LSP catches 90% of errors: Cursor’s Composer + LSP still faster for surgical edits; keep Cursor for that mode.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	4.5	Best terminal-autonomous + IDE hybrid; Routines are force multiplier
Ease of Use	4	cc CLI intuitive; Routines YAML learning curve (~1 hr)
Value	5	$100/mo credit pool covers heavy usage; VS Code free
Support	3	Discord only; Routines docs sparse; billing edge cases slow

---

This is part of our AI Coding Tool Evaluation series. See full setup guide: VS Code + Claude Code Hybrid: The 80% AI-Written Stack