Privacy Comparison

Privacy characteristics of major AI coding tools.

Comparison Matrix

Tool	Open Source	Privacy Mode	Training	Retention	Jurisdiction
Cursor	No	Yes	Opt-out	30d*	US
Zed	Yes	Yes (BYOK)	Never	None	US
Windsurf	No	Yes (ZDR)	Never	Zero	US/EU
Copilot	No	Yes	Opt-out	Varies	US
Continue.dev	Yes	100% local	Never	None	None
Claude Code	No	Enterprise	ON**	30d-5y	US
Aider	Yes	BYOK	Depends	Depends	None

*Non-Business plans
**Consumer accounts since August 2024

Truly Private Options

For maximum privacy, use local models:

Setup	Privacy	Quality
Continue.dev + Ollama	Maximum	85-90%
Zed + Ollama	Maximum	85-90%
OpenCode + Ollama	Maximum	85-90%

No data leaves your machine. Trade-off: requires capable local hardware (24GB+ VRAM for the best local coding models).

Hidden Gotchas

Tool	Gotcha
Cursor	30-day retention even with Privacy Mode (non-Business)
Claude Code	Training ON by default for consumer accounts
Anthropic	Must opt-out within 30 days or data used for training
Azure OpenAI	30-day abuse monitoring (waivable)

By Use Case

Personal Projects

Any tool is fine. Enable privacy mode if available.

Client Work (NDA)

• Safe: Continue.dev + local, Zed + BYOK
• Acceptable: Cursor Teams, Windsurf Enterprise
• Avoid: Claude Code (consumer), any tool with training enabled

Enterprise/Regulated

• Required: Cursor Enterprise, Windsurf Enterprise, or self-hosted
• Verify: DPA, subprocessors, audit logs, SSO/SCIM

Jurisdiction Notes

Location	Risks
US	CLOUD Act, FISA 702
EU	Generally stricter protections
China	PIPL, data localization requirements

For EU companies: Consider Schrems II implications with US tools.

Next Steps

• Privacy Deep Dive — technical details
• Security Risks — threats and mitigations