Launching today
Lykeion
A research lab where AI agents and researchers work together
1 follower
A research lab where AI agents and researchers work together
1 follower
Lykeion is an open-source research lab where AI agents carry out scientific work and people steer it. Each question becomes a study that moves through Grounding, Hypotheses, Experimentation and Validation, advancing only when its gate passes. Every claim an agent keeps cites a recorded source from OpenAlex, Europe PMC or arXiv, and review of each study is powered by Jev. Agents run on your own machine, inside a sandbox.






Free Options
Launch Team / Built With

AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted

Hi Product Hunt! I'm Diego, and I build Lykeion.
The name is an homage to the Lykeion of ancient Athens, where Aristotle founded his school: a community that gathered to investigate, argue and learn. It's remembered as one of the earliest research institutes.
The problem. Research today asks people to coordinate an enormous amount of work: finding evidence, developing hypotheses, running experiments and checking results. AI agents can take on more and more research work, but a pile of agents isn't a lab. They need a shared place to work, handoffs between stages, and rules for when a result is good enough to build on. And people need to see what was done, why, and on what evidence.
How it works
Each research question becomes a study that moves through four stages: Grounding, Hypotheses, Experimentation and Validation.
A study only advances when its gate passes. The server checks what it can verify mechanically, such as whether a cited source resolves. Review of each study is powered by Jev, which handles the judgment calls. Without a Jev key, you approve each stage yourself.
Every claim an agent keeps must cite a recorded source (OpenAlex, Europe PMC, arXiv), so any conclusion traces back to where it came from.
Lykeion doesn't ship a model. It drives the AI agents you already use, on your own machine, inside the OS sandbox. No sandbox, no run.
Does it help? In a blind test on 45 scientific claims from SciFact, I asked the same Claude model for quoted evidence, once on its own with web search and once inside Lykeion:
Quotes found word-for-word in the cited source: 52% alone, 82% in Lykeion
Cited sources that couldn't be reached: 22% alone, 2% in Lykeion
Cost for all 45 claims: $12.54 alone, $11.07 in Lykeion
It's a small test, but when Claude cites something inside Lykeion, it proves that the paper is there and the words cited are in it.
Thanks to Multica and Paperclip (github.com/paperclipai/paperclip), whose work on running AI agents together shaped how Lykeion is built.
Try it: the hosted lab, the desktop app for macOS and Linux, or self-host the whole stack with Docker Compose. It's open source under AGPL-3.0: github.com/lykeion-tech/lykeion. I'd love your feedback, especially on your first study: where did it get confusing, and what would make you trust its output?