Trust
Privacy & security
Where your prompt goes
| Surface | Where the engine runs | Prompt leaves your machine? |
|---|---|---|
Workspace (/app) | Your browser tab | No |
| CLI | Your machine | No |
| MCP server | Your machine (stdio) | No, but your MCP client may forward tool results to its model |
| VS Code extension | Inside VS Code | No, unless you choose Send to AI Chat |
| HTTP API | Cloudflare Workers | Yes, over TLS. It is processed in memory and never stored or logged |
What is stored
- No accounts and no server-side storage.
- The workspace keeps your draft, settings, and saved prompts in this browser's
localStorage. Clearing site data removes them. - If Cloudflare Web Analytics is enabled, it records cookieless page views. It never sees prompt text.
What is logged
For each API request: a request id, route, method, status, duration, an input-size bucket (such as 1k-10k), and whether it was rate limited. Prompt text and results are never logged.
Security model
- Strict Content Security Policy (hash-based scripts,
frame-ancestors 'none'), HSTS, andnosniff. - No third-party scripts except the optional Cloudflare analytics beacon. Fonts are self-hosted.
- The API validates every field, caps input size, rate-limits per IP, and returns typed errors without internal details.
- Sensitive-data detection returns categories and offsets, never the matched value.
- Prompt-injection detection is signature-based. It reports risk, not proof, and can miss novel attacks.
Report a vulnerability through a private GitHub security advisory on the repository.