Your open-source interactive agent to get things done

A little less hype, a little more substance. What about end-to-end ownership of GenAI flows? How can free open source enable a democratised tech landscape?

By Marius C

Oct 19, 2025

5 min read

A little less hype, a little more substance. What about end-to-end ownership of GenAI flows? How can free open source enable a democratised tech landscape?

The GenAI wave / the AI bubble ... isn't here, mate. Here is where the self-hosted open-source tools are. Tools. Not persons, not coaches. Not buddies. Tools. Hype is elsewhere. In fact, there is significant doubt that it is even a bubble. Rather than that, it seems if you don't bubble, you are in trouble.

Bubble, bubble, trouble, trouble

I find it alienating how much personality people project into select neural networks, which are used as pre-trained input-output systems. All these neural networks do is to infer. Predict. Something.

AI 'gents have no Intelligence of the W-questions you need to ask. Why, What, When, Where, ... Of course synthetic systems can infer. Some better than others, and that's the 2025-† trillon-dollar venture sector. But that † looks like it's rather 2040 than next year.

Substance

Generative "Artificial Intelligence" (AI) systems are tools that produce output even if the rules aren't clearly specified. Fascinating, if it works well. Enraging, if it doesn't. Highly dependent on the presets. Take a plain LLM like BitNet (which is great research work by Microsoft), and test it. The results will be underwhelming in comparison to the market leaders. So where is this edge coming from, that allows US big-tech to lay off so many workers?

A plain open-source LLM isn't a product. Plain LLMs are like little fish bubbling in the wide open sea. The "magic" that makes all of it seem like "AI" is

high-quality training data
...
specific high-quality training data (and this means curated labeled data)
prompts
model architecture
GPU hardware
...

No AI venture shares its entire training data. You cannot rebuild a model without this data. We are speaking of high-quality training data and a large enough quantity. That costs millions to curate / scrape / filter / enrich etc.

Components of an agent to assist you

Furthermore Agentic tools like claude-code or gemini-cli have large prompts. That's what holds these systems together, in combination with this corpus of training data and, of course, the model architecture. In the AI development world, often the training data isn't strictly legal [1] [2], and the model architecture isn't strictly unique [3]. We know where well how these models work and how to make them.

The pioneer and leader of Agentic assistants is Anthropic. Their claude-code has a GitHub repo, but it is not open-source. 🐟 - This is what has driven me to research alternatives. Which you can own end to end. Rebuild. And use with open-source models.

Open-source tools have advantages

flexible choices of providers and models
- In the past months, Anthropic has been using so-called quantized (down-scaled) variants of its own models. These models are less precise but cheaper to host. Great for Anthropic, but bad for the customers. Users see fluctuations in quality [4].
- I start to see many claude-code bugs regarding screen flickering and conversation history management.

Open-source agent: LLxprt-code is flexible

llxprt-code can use all kinds of models. And if you don't depend on a single provider, you can always switch.

LLxprt open-source features:

many providers (and not just API keys, also accounts), include Anthropic and Google. But also open models from Moonshot, or GPT-OSS via OpenRouter.
you can filter OpenRouter providers (by precision for example)
llxprt is a fork of gemini-cli
1. tool calling is robust
2. well designed
3. conservative use

Simply put, it's what claude-code is not. Open.

Too many terminal AI tools

One tool, few words. I don't know what the future will bring. I am ready. Furthermore, I'd rebuild models on hardware if I had the data. But sadly, in the West we have copyright laws. When I was working in China, I learned that they do not have that. That's a big advantage. DeepSeek, however, wasn't a side project by some quants. It is a statement From China with Love.