Privacy & Security

Defender scans every user prompt, every file read by the agent, and every command output before it leaves your machine. 18 detection patterns identify and tokenize:

Each detected value is replaced with a cryptographically random token (e.g., [EMAI-A1B2C3D4E5F67890]). The mapping is stored in an in-memory vault on your machine — never transmitted, never logged. Provider responses are detokenized locally before display. Streaming responses are handled safely: tokens spanning chunk boundaries are buffered until complete.

Defender collects zero data. No telemetry, no analytics, no crash reporting, no feature flags phoning home. There are no Defender servers. All code runs on your machine. Every line is open source (MIT license) and auditable.

Defender ships with a bundled provider registry — no network calls to models.dev or any external service at runtime. Works in air-gapped environments with local models (Ollama, LM Studio, vLLM). Self-host the provider registry with your organization's approved model list. Custom regex patterns can be added for organization-specific PII types.

Defender is MIT licensed. View on GitHub.