Introducing OpenScience. A better, open-source Claude Science. • Any model: GLM, Kimi, DeepSeek, Claude, GPT, your own fine-tune. Switching is one flag. • 250+ research skills across ML, comp bio, cheminformatics. All readable, editable, extensible. • No throttling, no gatekeeping, no one vendor deciding what science is okay. •... Native Atlas integration: many agents, one shared reproducible research graph. • Runs on your infra. Your data stays yours. Scientific AI should be open. One company shouldn't own the tools the rest of us discover with, or decide who gets to.show more

Open source. Star it, fork it, break it, build on it.

It's crazy - Fable is blocking any lookup of this because it talks about machine learning. Absolutely ludicrous @AnthropicAI @karpathy surely you know this is not the way?

For the 250+ built-in research skills (especially in chemistry and biology), are they executed within a secure, containerized sandbox by default, or does OpenScience expect the user to manage environment isolation on their own host infra?

Seems like a beefy project

this is awesome. science needs more open-source ai tools. "Runs on your infra. Your data stays yours" is exactly what scientific labs need to adopt more ai tools.

the strongest part of this launch isn’t another model integration. it’s the argument for keeping scientific research open from day one.

Open-source Claude Science is a pretty compelling framing. Vendor lock-in for scientific research never felt like the right direction.

This is what scientific AI should look like: open, transparent, and researcher-owned.

This is an excellent step towards more transparent science. AI tools especially the ones in EdTech should be highly interpretable so that we understand how an output is produced against a specific input.

For people already on a Codex plan, it shouldn't be too hard to use the subscription instead of an API key, right? Just ask codex to clone the repo and make any changes necessary?

Being able to switch between Claude, Kimi, GLM, and DeepSeek with a config change instead of a migration project is huge.

Fire

@ycombinator As long as it doesn’t block simple questions on genetic engineering

FOLLOWED!!!!! We rise together! •

The one-flag model switching across GLM Kimi DeepSeek Claude GPT and fine-tunes stands out. 250+ readable editable skills in ML comp bio and cheminformatics plus Atlas graph integration look solid for reproducible work.

Finally someone did this!

A great step for open, reproducible science.

This is the right direction for scientific AI. Research tools should not be locked behind one model, one vendor, or one moderation layer. The most important part here is not just “open source.” It is: • model choice • editable skills • reproducible research graphs • private infra • shared agent memory • no vendor lock-in Science needs AI systems that researchers can inspect, modify, and trust. OpenScience feels less like a Claude clone and more like research infrastructure. #OpenSource #AI #ScienceAI #AIAgents #Research #LLM #YC

damn pretty impressive, looking forward in contributing to it !

Open-source scientific AI is a big deal if research tools depend on one vendor, science gets rate limits, pricing changes, and permission walls. fr, discovery should not need a subscription boss

So sick

one flag switches the model, not whether the skill actually works. i've had prompts that were bulletproof on claude turn into slop on smaller models with zero changes to the prompt itself

Check dm @SynScience

Hi! Please get in touch with me, I’ve been in this field for a while and looking to grow and collaborate. I currently offer training on how to Fine Tun.

im not sure how relevant this is to people in the comments but I (*coughs* fable5) started working towards a universal theory of physics:

Love hearing the words Open-Source and AI together, It's our favorite thing

i'd use this for the atlas graph alone

I LOVE THIS REVOLUTION 😀 FUCK CLAUDE

the boring part is not the model menu. it is convincing labs that "open" also means exportable without filling six institutional forms.

The "one flag to switch models" pitch is genuinely compelling, but that's also where I'd slow down before celebrating. Swappable models sound like pure flexibility until you're in a clinical or research context where the choice of model isn't just a preference, it's an auditable decision with liability attached. And that's the tension I kept running into when writing about Llama Stack's architecture at where the standardized API layer matters precisely because it separates infrastructure from model behavior in ways that can be validated separately. Open source code you can read isn't the same as open source you can audit for a specific clinical decision. The 250+ research skills being readable and editable is the real story here. But editable by whom, under what governance, with what validation trail? The no-throttling point cuts both ways too. Throttling is annoying for researchers. In a regulated pipeline it's sometimes the mechanism that forces human review at volume thresholds. None of this is a knock on the project. The extensibility model is exactly right for comp bio and cheminformatics where every lab has idiosyncratic data formats. But the governance layer is where open source health and science tools either earn institutional trust or stay perpetually in the "promising demo" category.

skill suggestion: bypass-paywalls. give any paper, and use all means to retrieve text

We really got open source Claude Science 5 days after Claude Science

finally!