PruneJuice: Building Privacy-First AI Tools in Your Browser

I’ve been working on something that started with a practical problem: I have access to several AI models – Claude Pro, Gemini Pro, and GitHub Copilot – and wanted to move conversations between them in a simple and direct way. That’s how PruneJuice was born. It’s here: https://prunejuice.me

The Origin Story

Claude would be brilliant at working through complex code architecture, then hit its context limit. Gemini would excel at data analysis that Claude struggled with. ChatGPT via VSCode had its own strengths. But there was no good way to port a conversation from one to another without just doing ‘select all’ and copy and paste. This seemed to me to be a huge waste of tokens – but the alternative of scrolling through a long conversation and curating just the bits you wanted to use was also boring and time consuming – and I didn’t want to ~pay~ for yet another web service that I would only use occasionally.

Copy-pasting chat history is tedious and loses structure. Export formats can be confounding and unhelpful. Most importantly, I didn’t fancy sending my conversation history through someone else’s server just to reformat it.

So I built PruneJuice – a tool that lets you compress, prune, and export AI chat data. The critical requirement: it runs entirely in the browser. Your conversations stay on your machine, processed locally. Nothing leaves your device.

The compression algorithms are very simple – whitespace removal and simple dictionary substitution ciphers. It can produce human-readable or non-human-readable text – LLMs have no problem at all decoding what seems like an incomprehensible wall of text – if they do, you can just accompany it with the dictionary, which you can also download from the site. It’ll give you some rough idea of the compression – but the main use is that it’s super simple for when you run out of tokens in one LLM – to carry on where you left off in another, without skipping a beat. It’s much more cost effective than ‘upgrade now to super-expensive ultra tier’ on Model X – plus you get the benefits of multiple LLMs working on whatever problem you have in mind.

Why Client-Side Matters

When working with AI chats, you’re often discussing code, business, creative plans, and half-formed ideas. PruneJuice runs entirely in your browser – no backend, no database, no data collection. Just HTML, vanilla JavaScript, and CSS.

The security model is straightforward: if data never leaves your machine, it can’t be leaked. The entire application is a single HTML file you can inspect yourself. Plus – I don’t have to host any complex back-end stuff – you get the webpage and just run it. It’s nothing to do with me – no guarantees and no idea what users do with it. Even the stats on it at the bottom of the page are local to the user – I can’t access them and I’m not interested. Mind you – the usual website caveats apply – this is hosted on cpanel, so all the usual webserver access logs are probably on there somewhere. But that’s only about browser agent, whatever geoIP stuff is passed – nothing about the content that users use the website for. There’s nothing sinister about that – just that it’s not my interest or responsibility if anyone chooses to use the tool for any purpose.

From Compression to Workflows

Once the compression tool was working, it became clear to me that I wasn’t just having linear conversations. I build workflows – RAG systems, prompt chains, pipelines where output from one model feeds into another.

That led to PruneJuice Flow – a visual programming environment where you drag and drop nodes to build AI prompt workflows. Pull content from Google Drive, feed it to Claude, chain it into another prompt. Process images through local Ollama models. The idea was to make experimenting with these AI workflow patterns straightforward.

I’m not sure if some of it even works properly yet! – I mainly use it for creating JSON prompts for generative AI systems like comfyUI – but I’ll get round to testing it properly when the need arises. So please do accept it more as a sketch rather than a polished tool with guaranteed workability.

Like many a self-employed solo dev, I do this entirely on my own time – so if you find it useful, feel free to make a small donation (I think you can add a note to that) – and I can get an idea of how to improve functionality and whether it’s actually working or not. I don’t expect to sell any dried prunes through the Amazon associates program.

Technical Approach

The technology choices were deliberately simple: vanilla JavaScript that runs in any browser, no React, no build pipeline, no npm dependencies. State lives in memory – canvas layout, nodes, connections. Save to a local JSON file, load by dropping it on the canvas. I do actually have a snazzy react version coming at some point – I’ve lost interest in it for the time being, but will probably revisit it some time in 2026/7 to see if it’s worth the effort.

localStorage handles preferences only. Blob URLs handle media from models. Google Drive integration uses OAuth 2.0 directly between your browser and Google – well, it will eventually, I don’t think that’s working at the moment as I haven’t needed it yet.

Current State

Both tools are running live at prunejuice.me – in a functional-enough but incomplete state. PruneJuice Flow is probably best for in-browser JSON prompt flows using the visual canvas, with an unfinished plan to support any OpenAI-compatible API (including local models like Ollama), Google Drive integration for RAG workflows etc.

The interface could be more polished. Error handling could be better. But it works, and it does what I needed it to do: let you move conversations between models and build workflows without sending your data through external servers.

I’m planning to open-source it at some point when I have a moment. These kinds of tools should be freely available and inspectable – and the best way to demonstrate there’s no data collection is to show the entire codebase.

Try It

The tools are available at prunejuice.me if you’d like to experiment. They’re free to use, run entirely in your browser, and might be useful if you find yourself juggling conversations across different AI models. Use entirely at your own discretion.