Transparent proxy strips credentials before LLM requests

DontFeedTheAI is a transparent proxy that strips IPs credentials hostnames and PII from every request before it reaches the AI and restores them on the way back. The system uses a two-layer pipeline: a local Ollama layer that detects sensitive content in prose and a regex layer that catches IPs hashes tokens API keys.

Both the Ollama detection and regex rules run on the user's machine. No sensitive data crosses the boundary to external services. The proxy works with Claude Code OpenAI SDK OpenRouter Requesty or any OpenAI-compatible client.

Use cases across security and development roles

  • Pentesters can run nmap mimikatz bloodhound output through Claude without exposing client infrastructure
  • Developers and SREs can debug with production data or internal configs in regulated environments
  • Legal and consulting teams can anonymize client contracts case files or proprietary IP in AI-assisted reviews
  • Finance and compliance can analyze reports or audit scripts without exposing account details
  • Researchers can query LLMs on confidential datasets

Architecture compared to alternative approaches

Cloud anonymization API plus LLM creates two bills and sends data through more hands. Ollama alone leaves data on the machine but has no awareness of what is sensitive. Claude or OpenAI directly offers best reasoning quality but everything lands in their infrastructure with real client IPs credentials org names in API logs. DontFeedTheAI provides cloud reasoning quality with local detection and nothing sensitive crossing the boundary.

Installation and operation

With a VPS recommended for team use or persistent engagements: clone the repository run the wizard which asks engagement name VPS address model then deploys opens the SSH tunnel and launches Claude with the proxy active. Locally without a VPS: create a virtual environment install dependencies start proxy and Ollama in Docker set environment variables for ANTHROPIC_BASE_URL and ENGAGEMENT_ID then launch any OpenAI-compatible client. The system works on Windows macOS and Linux.

Validation tools included

Two tools ship with the proxy to help validate coverage and extend it. Visual audit open in browser while the proxy is running shows every original to surrogate mapping logged during the session filterable by entity type with per-request timing breakdown. Testing the full pipeline runs all 53 fixtures through the complete LLM and regex pipeline asserting zero leaks. Auto-improvement loop tightens the regex floor reporting leaks and false positives and tells exactly which strings slipped through contributing cycle adds a fixture for a real tool runs the loop adds a regex pattern for each leak repeats.

Founder's rationale

This was not built to be innovative there are already cloud APIs that do LLM-based anonymization but that means sending data to yet another third party the author refuses. If you work in security you already know why. The architecture is made available to everyone and the community can help expand its effectiveness for free you are paying for context processing the AI does not need your real data for that.

— zeroc00I