sgasser

Everyone says don't send personal data to cloud LLMs. But when you're working with customer emails, support tickets, or code with credentials — it's hard to avoid.

So I built a proxy that handles it for you — it's open source and free.

How it works:

  You send:        "Email [email protected] about meeting Sarah Miller"
  LLM receives:    "Email [[EMAIL_1]] about meeting [[PERSON_1]]"
  LLM responds:    "Dear [[PERSON_1]], I wanted to follow up..."
  You get back:    "Dear Sarah Miller, I wanted to follow up..."

PasteGuard finds personal data and secrets in your prompt, swaps them with placeholders, and restores the real values in the response. The LLM never sees the actual data.

What it catches: - PII: Names, emails, phones, credit cards, IBANs, IPs (24 languages) - Secrets: API keys (OpenAI, Anthropic, AWS, GitHub), JWTs, SSH keys

Works with both OpenAI and Anthropic APIs. Point your app to localhost:3000/openai/v1 or localhost:3000/anthropic/v1. Compatible with LangChain, Cursor, Claude Code, Open WebUI.

One command to run:

  docker run -p 3000:3000 ghcr.io/sgasser/pasteguard:en

Happy to answer questions.

Using LLM APIs but worried about sending client data? Built a proxy for that.

OpenAI-compatible proxy that masks personal data and secrets before sending to your provider.

Mask Mode (default):

  You send:      "Email [email protected] about meeting Dr. Miller"
  LLM receives:  "Email <EMAIL_1> about meeting <PERSON_1>"
  You get back:  Original names restored in response

Route Mode (if you run a local LLM):

  Requests with PII  →  Local LLM
  Everything else    →  Cloud

What it catches:

  PII: Names, emails, phones, credit cards, IBANs, IPs, locations (24 languages)
  Secrets: Private keys, API keys (OpenAI, AWS, GitHub), JWT tokens

Uses Microsoft Presidio for PII detection. ~500MB RAM, 10-50ms per request.

Works with Cursor, Open WebUI, LangChain, or any OpenAI-compatible client.

Docs: https://pasteguard.com/docs

Feedback on edge cases welcome.

Thx rentalsunited looks great.

Thanks. :)

Not at the moment. Maybe Airbnb in the near future. Any other suggestions?

Thank you. I am very pleased to hear that.

Thx!

For the first release we analyze the popularity of an article by likes, shares, count of reads, etc. For the second step we analyze also the body.

We are using only RSS, so we have no header/footer/comments/navigation. Ads we remove by converting html to text.

I am working on a online newsreader that allows you to quickly read only important articles from websites and blogs of your choice. Every article gets analyzed and evaluated, and uninteresting articles get filtered out. So you don’t have to read all articles of your favorite websites and blogs – only the important ones will be shown to you.

http://zebrello.instapage.com