I Rebuilt My AI Research Swarm to Run 100% on My Laptop!

A few weeks ago I built a research agent swarm with MiroFish, and the big problem was it leaned on cloud components to get running. So I pulled the research side out and rebuilt it from the ground up as its own product, Liberati Research. In this video I demo the whole thing running on my MacBook, using open source components, free news feeds and a local model. No accounts, no API bills, no middleman!

What does Liberati Research actually do?

You type in a question and it goes to work. The AI looks at your question, figures out which news sources are best for that topic, and pulls the articles straight onto your computer. Just under 10 news sources are wired in right now, and a lot of them have free APIs you don't need to sign up for.

During testing I've pulled in around 5,000 news articles, 300 markets and some PDF files, all for free. The markets side pulls in the top Polymarket questions. You start with the question, gather the news, then work out the most likely outcome.

All those articles on a topic become a corpus. I also ingest the corpus into a Postgres database so there's a RAG chatbot at the end that you can question about the data set.

How do the agent debates work?

For the demo I asked a macro question: who is winning on market share between Tesla and BYD? The pipeline looks like this:

1. Retrieve all the news data from the sources onto the local machine
2. Create four agent personas from the question
3. Round one, each persona reads the corpus and forms an answer
4. Round two, they debate, and sometimes take the opposite side to the previous round
5. Synthesis, a report brief at the end
6. Visualisations of the whole lot

The brains are Hermes 3 running through Ollama on my MacBook, which I find is really fast. It did heat the laptop up a bit 😉

The four personas were a global electric ecosystem angle, a geopolitics angle, a BYD company angle and a Tesla angle. Honestly their names came out like magazine titles ("BYD's Battery Boost", "Elon's Electric Edge") rather than human personas, so that needs work. The synthesis said BYD has topped Tesla as the biggest EV seller in a calendar year. It wasn't specific on the year, I'm assuming 2025. The agents also disagreed on parts of it, mainly about comparing Tesla's growth rate to another company's, which wasn't the quantitative comparison they wanted.

This is also a completely different UX to what I showed a few weeks ago. Previously you started with a globe and clicked through topics. From my testing the best way to research something is to start with the question and let the tool do the research and categorisation for you.

What visualisations do you get?

• A live auto research button, inspired by Karpathy's auto research, where the news keeps growing and the agents keep refining their opinions
• Corpus analysis that tries to map articles left or right and show whether they repeat the same story or give different opinions (this section needs more work)
• A global map plotting where articles came from. For this question most were from China, with one from Britain
• A swarm mode where five agents keep reviewing the data and ask each other questions, more like agents on Reddit than a structured debate
• A news source timeline and a full list of the articles so you can click through and check they're legit
• A "next action" step. In my case that was drafting a YouTube script on BYD vs Tesla, but it could be a PDF report or a presentation

Then I asked a follow up about the long term trend, whether BYD will dominate on charging infra and battery tech. The local model streamed back an answer quoting the articles it had fetched, including BYD at 15.7% market share and its lead in battery tech and charging, while noting the EV market is intensely competitive.

Why sovereign?

The main thing is all of this happened on my computer. I own the whole tech stack and you can read all the code. The news research is done in Python against free providers, the models run through Ollama, and once the feeds are pulled in you can keep doing analysis without the internet.

You get the repo through the Liberati school on Skool. It's the cost of a $4 cup of coffee a month, with premium tiers if you want extra help and blueprints. If there's an open source AI tool you want built, drop it in the community and we'll build it together.

Stay sovereign! Join us at liberati.life.