sure-scale/hey-jude

Hey Jude is an open-source proxy that sits in front of large language model calls and tries to reduce the amount of sensitive information those models ever see. It is aimed at legal workflows, where prompts often contain client names, organization names, email addresses, and phone numbers that should not be shipped to a third-party model without first being scrubbed. The README is clear that this is a helper layer for data minimization, not a guarantee of complete detection, and it should be used alongside other confidentiality measures. The gateway works by spotting common personal and organizational details in your prompt, swapping them with safer synthetic values, sending the cleaned-up prompt onward to the model you actually want to use, and then attempting to restore the original details in the response that comes back. The pseudonymization step itself is done by a local model, usually run through Ollama, so the raw text never has to leave your machine just to be cleaned. It exposes OpenAI, Anthropic, and Gemini compatible endpoints, which means existing client code written for those SDKs can be pointed at the gateway with very few changes. The README shows short Python and Node examples for both the OpenAI and Anthropic SDKs, where the only real change is the base URL and the API key. The default setup is meant to run end-to-end on your own laptop without any cloud API key. You install Ollama, pull a small Qwen model, clone the repo, copy the example env file, and start it with Docker Compose. Redis runs in the Docker stack and Ollama runs on the host. Once it is up, a test script confirms the gateway is reachable on port 4005. Advanced users can swap in larger local models, switch to Apple MLX, or route the cleaned prompt to OpenAI, Anthropic, Gemini, Azure, or any LiteLLM-compatible provider by setting the right environment variables. The project is released under the AGPL-3.0 license.