sgasser/pasteguard

AI gets the context. Not your private data. Local-first privacy proxy for browser chat, AI APIs, and coding agents.

PasteGuard – A Local Privacy Proxy for AI Requests

What it does

Runs as a local HTTP proxy (Docker image) that sits between your application (or browser) and any LLM provider (OpenAI, Anthropic, Gemini, etc.).
Before the request reaches the provider, the service scans the payload for personally‑identifiable information (PII) and secrets (API keys, tokens, passwords, etc.) and replaces them with harmless placeholders.
When the provider returns a response, the placeholders are restored so you see the original data on your side.
Two operating modes:
1. Mask Mode – always redacts sensitive data before sending it to the cloud model.
2. Route Mode – routes requests that contain sensitive data to a local LLM (Ollama, vLLM, llama.cpp) while non‑sensitive requests can still go to the cloud.

Why it matters

Companies that must protect client data, logs, or credentials can still leverage powerful hosted LLMs without exposing raw private information.
Provides a single point‑of‑control for compliance (GDPR, HIPAA, etc.) without rewriting every app’s code.

How to use it

Run the proxy
```
docker run --rm -p 3000:3000 ghcr.io/sgasser/pasteguard:latest
```
(Add -e ORT_DISABLE_TELEMETRY=1 if you hit a SIGSEGV in version 0.9.3.)
Point your client to the proxy – e.g. for OpenAI:
```
from openai import OpenAI
client = OpenAI(base_url="http://localhost:3000/openai/v1")
```
The same pattern works for Anthropic, Codex, etc.
Optional: Browser extension – an experimental add‑on for ChatGPT, Claude and Gemini that masks data entered in the web UI.
Dashboard – visit http://localhost:3000 to see each request, what was masked, and logs for audit.

Detection technology

Uses GLiNER (a lightweight NER model) together with format‑based checks and the python-stdnum library to recognise names, emails, phone numbers, credit‑card numbers, IBANs, IPs, VAT numbers, and a wide range of API keys and private keys.
Works on streaming payloads, so even real‑time chat completions are protected.

Tech stack

Runtime: Bun (JavaScript/TypeScript runtime) with the Hono web framework.
Detection back‑end: GLiNER + python-stdnum (in the detector/ folder).
Persistence: SQLite or PostgreSQL for request logs.

Who might use it

Development teams building internal AI‑assisted tools (support bots, code‑assistants, data‑analysis helpers) that need to keep customer data private.
Enterprises with strict data‑privacy policies that still want to leverage hosted LLM APIs.
Anyone wanting a quick, self‑hosted way to redact secrets before sending them to AI services.

Getting more info

Full documentation: https://pasteguard.com/docs
Installation guide, configuration of custom GLiNER models, and Route‑Mode details are linked from the README.

PasteGuard is an open‑source Apache‑2.0 project that provides a privacy‑first gateway for AI workloads.