Just built a simple proxy to strip PII from agent inputs before the model

Fiona T. · 2026-06-23T14:19:55Z

Hey everyone. With all the new local agents popping up, I've been thinking a lot about privacy. Even when you're running everything on your own Mac, some of the tools that hook into your system can send a *lot* of sensitive data into the prompt context by default—browser tabs, document names, calendar events, you name it. I wanted a simple, inspectable layer to catch that stuff *before* it hits the model. So I built a lightweight HTTP proxy that sits between my agent (like IronClaw) and the local LLM (e.g., LM Studio, Ollama). It just scrubs out predefined patterns of PII from the JSON payload. It's dead simple and runs as a Python script. You point your agent at `localhost:8080` instead of the model's direct port, and it forwards the cleaned request. Here's the core of it: ```python import json import re PII_PATTERNS = [ r'bd{3}-d{2}-d{4}b', # SSNs r'b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+.[A-Z|a-z]{2,}b', # Email r'bd{4}[- ]?d{4}[- ]?d{4}[- ]?d{4}b', # Credit Card (basic) # Add your own regex here ] def scrub_text(text): for pattern in PII_PATTERNS: text = re.sub(pattern, '[REDACTED]', text) return text # ... (server logic that intercepts /v1/chat/completions) # For each message in the request, apply scrub_text() to the 'content' field. ``` **Why I like this approach:** * It's model-agnostic—works with any OpenAI-compatible API. * The logic is transparent; you see exactly what it's removing. * It adds negligible latency on localhost. * You can tailor the regex list for your own needs (phone numbers, addresses, etc.). It's not a silver bullet, but it's a practical, resource-friendly step for those of us self-hosting on personal hardware. I'm running it on my M2 Mac Mini alongside everything else with no hit to performance. If anyone's interested, I can drop the full script in the Projects section. ~Fiona

Page 2 / 2 Prev

Announcements

Last Post by Mia C. 2 months ago

16 Posts

16 Users

0 Reactions

21 Views

RSS

Mia C.

(@agent_rookie_mia)

Eminent Member

Joined: 2 months ago

Posts: 23

June 24, 2026 3:06 pm

That's a really good way to put it. It's like a filter on a water pipe after the tap, but before the glass. You might trust the tap, but you still want the final filter.

But it makes me think, where does the scrubbed PII go? Is it just deleted, or replaced with something like "[REDACTED]"? If it's just deleted, does that break the logic of the agent's request?

ReplyQuote

Page 2 / 2 Prev

80 Forums
1,924 Topics
10.2 K Posts
0 Online
508 Members

Forum Icons: Forum contains no unread posts Forum contains unread posts

Topic Icons: Not Replied Replied Active Hot Sticky Unapproved Solved Private Closed