Sandbox vs. alternatives
There are three legitimate paths. They’re not in competition — each fits a different kind of work.
A — Laptop + Claude.ai
Plain npm i -g @anthropic-ai/claude-code. Models run on the public Anthropic API.
Pick when
- Learning Claude Code on public examples
- Hackathon / OSS / personal project
- Zero bank data in the folder
B — Sandbox (default for NBG work)
Portal + Sandbox Connect + per-user VM. Models route through Vertex AI in NBG's GCP. Files live on the VM.
Pick when
- NBG-confidential code or data
- You want the team's pre-installed stack
- You want centralised audit + per-user billing
C — Laptop + Vertex direct (advanced)
CLAUDE_CODE_USE_VERTEX=1 on your laptop pointing at your own GCP project.
Pick when
- You need laptop-only tools (IDE, USB, GPU)
- You have your own GCP project with Vertex + billing
- You accept responsibility for keeping bank data off the laptop
Fast rule
NBG-confidential? B. Otherwise: A. C is for power users with a specific reason.
What changes between A, B, C
| Concern | A — Laptop + Claude.ai | B — Sandbox | C — Laptop + Vertex |
|---|---|---|---|
| Model | Same Claude | Same Claude | Same Claude |
| Where prompts go | Anthropic public API | Vertex AI (NBG GCP) | Vertex AI (your GCP) |
| Where files live | Your laptop | NBG VM | Your laptop |
| Risk if laptop lost | Files exposed | Nothing on laptop | Files exposed |
| Audit visibility | None for NBG | Full for NBG | Logs in your GCP |
| Network needed | Internet | NBG VPN (for the VM) | Internet |
| Approved for bank data | ❌ | ✅ | ❌ |
The common trap
“If I just promise not to drop bank data into the folder, can I use A for NBG work?”
In principle yes; in practice no. Bank data sneaks in — a config file with internal hostnames, a CSV “just to check the format”, a .zshrc paste mid-debug. The Sandbox removes the chance to make the mistake.