GitInTheVan allows you to attach lorebooks/scripts to your personas, pass responses through a secondary LLM, etc.

Why make this?

With my decision to step away from Janitor, I have returned to a previous project I was developing.

This was originally going to be my personal alternative to Lorebary after I saw they went from being open-source (Even with a limited license, being able to see the code to verify it wasn't doing anything malicious was enough for me to trust them) to closed source.

I have no problem with Lorebary's creator or its staff. I love the concept, I simply hate the execution.

The long and short of it is that it's yet another black box application trying to do too much.

1. They have their own LLM

2. They have their own user profiles

3. They have their own comments

In my opinion, most people using these platforms already have profiles on JanitorAI and other platforms.

Stacking all these features on top of a site that's already heavily strained and vibe-coded is just begging for constant performance issues. I'm running on a Ryzen 9, 64GB DDR5, and a fiber connection. The site runs at like 5-10 FPS. That's without logging in or taking any other action.

What makes it different?

1. It is focused on 'Cantrips', what I'm calling javascript-enabled lorebooks. These are not Scripts, because while they will be backward compatible with JanitorAI's scripted lorebooks, these will also have access to persistent memory. You can also control if they fire before or after the LLM takes its turn.

2. It includes LLM-powere validation. You can set rules that trigger AFTER an LLM fires off its reply. The reply is then processed by the LLM while reviewing the rules, and the LLM either marks it as a pass or fail. If it fails, it will send the original prompt and additional instructions for a retry. It will do this 'n' times, however many times you specify.

3. It allows LLMs to take multiple actions using Cantrips. On JanitorAI (And Wyvern, I believe), scripted lorebooks happen between the user and the LLM's turn. GitInTheVan allows you to do something like setting up my dice script, letting the LLM decide it needs to roll dice, and then the LLM can receive the results so it can finish its turn.

4. It is fully self-hosted and fully Open-Source.

Lorebary is fundamentally an input-side prompting and injection engine. GitInTheVan will be capable of these features but it is primarily a capability extension and validation engine.

What can it do currently?

1. Cantrips (Backwards compatible with JanitorAI Scripts, running in a sandbox for safety) and Lorebooks

2. Tagged injection, i.e. put <#cantrip-tag#> in your persona and it'll load that cantrip, same with lorebooks

3. Verification - You can attach rules and use a different LLM endpoint for your verification bot

4. Runnable on Linux, Windows, and Mac OS

5. Has some basic Cantrip templates

6. Auto summarizes, retaining a specific number of full messages sent to the LLM and summarizing the rest

7. I'm actively working on this today so this list may change dramatically

What will it have in the future?

I'm doing my best to avoid feature creep, but in the future it will:

1. Allow your LLMs to take turns for rolling dice and reading rules. The initial plan is to inject instructions and a summary of available options. In the future I will be adding an MCP, RAG, and a skill library.

2. Instead of acting as a social media site with profiles, etc., I am going the SillyTavern route where if you want to share your Cantrips and lorebooks you would share them via git repos (Github, etc). Since Github blocks explicit material I'll be allowing any standard git but we'll include sandboxing to ensure files are safe.

3. Options to chain verification LLMs. Example: You have a 1B parameter model that does nothing but verify outfits haven't changed? You can insert it and have specific rules for it. You want to run your verification by a locally hosted 12B model before you send it to a larger model (Which would weed out false positives)? Sure, go for it.

4. Options to chain writing LLMs. Gemma 4 31B IT is great at following instructions for adhering to markdown formats (Asterisks for actions, for examples) but its prose can be a bit dry. You can have one model draft, feed it through verification, then feed it through a second model for prose. Or do the reverse - Write the scene, then apply the formatting.

5. Optionally save your full chat history instead of simply retaining a summary/memory. Because this is self-hosted, this protects you from being banned by ensuring you retain everything.

6. There is more, but a lot of the more complicated things will be later and dependent upon my ADHD to remain focused.

What is needed for testing?

Try it. Give me concrete feedback on the UI, the install documentation, any testing you do.

Use it with various providers as your endpoint. I use Open WebUI + Llama-Swap at home, and so I've not yet tested it against other OpenAI compatible endpoints.

Report issues and provide feedback on Github or Discord. Note for Discord users: Send me a message with any friend requests letting me know you're coming from here so I don't assume you're a spam bot.

Requirements

Requirements are based on Vite + TypeScript + Svelte

Disk Space: 1GB (My install is at ~350MB)

RAM: 2GB RAM, 4GB recommended

CPU: Vite allows for pretty old stuff, this will run fine on most things

Operating System: MacOS, Windows, Linux

OpenAI Compatible LLM: Anything with an OpenAI compatible endpoint can be used, whether it's a self-hosted llama.cpp, Open WebUI, Chutes, OpenRouter.

Note to Moderators: This is an open-source self-hosted application. It is not a competitor to JanitorAI. As such it should be within the rules to post just as my Firefox plugins and other items, just like dozens of users have posted information about Sophia's Lorebary. If this is not correct, I apologize.

Disclaimer: This is open source and relies on upstream libraries in Python. I can promise that I'm not doing anything malicious, but there is always a risk of a 'supply chain attack' where an upstream library is infected. Know that if you're infected by something upstream, I'm probably cursing myself because I'll also be infected. Be careful and stay safe. Scan anything you download, and feel free to have an LLM like Gemini review the Github repo for any concerns.

Warning: I have been using an LLM for testing automation and validation while I've been doing the UI testing and validation myself. Just like my Scripts, my time for testing things is extremely limited due to family, work, etc. If you choose to test, you will almost certainly find bugs or UI issues. I would not use this assuming it will be a perfect product.