100% free, open-source alternative to Claude Science, use whichever you prefer between DeepSeek and GLM.
Academic researchers are thrilled!!
Less than a week after Anthropic launched Claude Science, the open-source community has delivered its own alternative solution.
An AI research team incubated by YC has released OpenScience, a fully open-source equivalent of Claude Science.
It is also a full-process AI research workbench covering literature retrieval, hypothesis generation, code experimentation, and paper writing, but it is not tied to any single model vendor.
DeepSeek, GLM, Claude, GPT… no matter domestic or international, you can use whichever model you prefer.
Moreover, the project adopts the Apache 2.0 license, the most developer-friendly agreement, and can be installed with just one line of command.
As soon as the news came out, the project immediately became a trending topic on X. People commented enthusiastically:
This is exactly what a scientific AI should look like. (Anthropic: Just mention our name directly)
Claude Science is powerful, but it's not accessible to many…
About 5 days ago, Anthropic officially launched Claude Science at a closed-door event hosted by MIT Technology Review.
This is an AI work platform designed exclusively for scientists, integrating various tools and software packages most commonly used by researchers.
For example, in the past, a researcher completing a study had to search for literature on PubMed, write code in Jupyter, run statistical analysis with R, connect to clusters via SSH to submit tasks, and use various tools to create charts and write papers.
Switching between a dozen windows constantly, the effort spent on transitioning between different tools alone consumed a huge amount of energy.
What Claude Science aims to do is integrate all these functions into a single workbench.
Specifically, it has implemented several key integrations:
At the database and toolchain level, it comes with over 60 pre-built scientific database connectors and pre-configured skill packages, covering common research fields such as genomics, single-cell analysis, proteomics, structural biology, and chemoinformatics.
When you ask questions in natural language, the specialized Agent will automatically perform cross-database queries, so you don't need to browse databases like UniProt, PDB, Ensembl, ChEMBL, and GEO one by one.
It also integrates NVIDIA's BioNeMo Agent Toolkit, enabling direct connections to life science models such as Evo 2, Boltz-2, and OpenFold3.
At the execution level, it introduces a multi-agent architecture.
The main Agent is responsible for overall planning, while sub-agents process different tasks in parallel. There is also a dedicated Reviewer Agent specifically for fact-checking, such as verifying citations, validating calculation results, and marking potential errors.
The generated outputs are not limited to plain text — 3D protein structures, genome browser tracks, and chemical structural formulas can all be natively rendered.
Moreover, every chart retains its generation code, runtime environment, natural language description, and complete conversation history simultaneously.
In certain scenarios, scientists can even modify charts directly with a single sentence, and the system will automatically rewrite the underlying code.
At the computing power level, Claude Science can directly connect to the existing infrastructure in your lab.
Laptops, Linux servers, and HPC cluster login nodes are all supported. You can call cloud GPUs on demand via SSH connection or a Modal account, scaling from a single GPU to hundreds of GPUs.
Large datasets only need to be loaded once, and sensitive data does not need to leave your own system — only the context required for each step of analysis is sent to Claude.
Early beta users have already achieved practical results with it.
Jérôme Lecoq, a neuroscientist at the Allen Institute, used it to build a multi-agent "computational review template" containing around 20 custom skills, allowing sub-agents to read thousands of papers, extract core viewpoints and quantitative data, and then generate reviews chapter by chapter.
To put it in perspective, writing a single review used to take two years, but now he has nearly 10 drafts in progress —
many of which are over 100 pages long, with all citations verified by the Reviewer Agent.
Stephen Francis from the UCSF Brain Tumor Center used it to conduct molecular epidemiological research on glioma and run germline variant analysis.
He said Claude Science reduced the time required to one-tenth of the original, and his team independently verified the results, confirming that the analysis is both fast and reliable.
Combined with the evaluation of AI research capabilities by Harvard physicist Matthew Schwartz in March this year, the current level of Claude is roughly equivalent to a second-year graduate student.
He published a guest article on Anthropic's official blog titled "Vibe Physics: The AI Grad Student", which documented his entire process of using Claude Opus 4.5 to complete a theoretical physics paper.
The conclusion he drew at that time was:
The current research capability of AI is roughly equivalent to a second-year graduate student: it can work continuously without complaining, but needs supervision from a mentor at every step.
This judgment was later incorporated by Anthropic into Claude Science's technical documentation as a calibration point for the product's positioning.
However, Claude Science currently has several hard limitations:
It only supports macOS and Linux
It is exclusively available to paid Pro/Max/Team/Enterprise users
Only Anthropic's own Claude models can be used on the platform
With all these barriers combined, especially for domestic research teams, Claude Science has become something that is "hard to reach despite being appealing".
Good News: The Open-Source Alternative Has Arrived
Aiming at the above limitations, the open-source project OpenScience came into being.
The team behind it is called Synthetic Sciences, founded in San Francisco in 2025, and just graduated from YC's 2026 Winter Batch this year.
The founding team has great ambitions: they want to build a platform where scientists can directly delegate complex research tasks to "AI co-scientists", allowing the AI to autonomously complete the entire workflow from literature review, hypothesis generation, experiment execution to paper writing.
They have a core internal judgment:
Scientific foundation models need to possess genuine "research taste", which cannot be achieved simply by stacking parameters. It requires parallel development of both products and models: using the product to collect high-quality research process data, and then training tasteful models with this data.
OpenScience is the first product implemented under this development roadmap.
Although OpenScience shares the same mission as Claude Science, there is a fundamental difference between them:
It is model-agnostic.
In Synthetic Sciences' own words:
Scientific AI should be open. No single company should monopolize the tools humanity uses for exploration and discovery, let alone decide who is eligible to use them.
Therefore, on this platform, models from Anthropic, OpenAI, Google, DeepSeek, GLM… as long as you have the corresponding API Key, you can connect to them directly.
You can even run local models deployed via Ollama, ensuring that no data leaves your machine at all.
Your API Key remains stored locally, and requests connect directly to the model provider without passing through any intermediate servers.
Furthermore, OpenScience supports switching models per individual request.
Within the same workbench, you can use Claude for this step and switch to DeepSeek for the next step, without modifying any configuration.
In terms of functionality, OpenScience is even more aggressive than Claude Science —
It has over 250 built-in research skill packages, more than four times that of Claude Science, covering fields such as machine learning, computational biology, and chemoinformatics, all of which are readable, editable, and extensible.
Installation is also very simple: just one line of command in the terminal:
It is ready to use immediately, and the workbench will automatically pop up in your browser. On first run, select a model source and enter your API Key, then you can start working.
If you prefer a global installation:
If configuring multiple API Keys feels troublesome, the team also provides a hosted platform called Atlas —
You can top up your wallet to directly call multiple cutting-edge models without configuring each API Key individually, and you also get access to persistent research graphs and cloud computing power.
However, Atlas is not mandatory. You can run OpenScience completely for free with your own API Keys, with no access barriers at all.
One More Thing
Interestingly, if you scroll to the very bottom of OpenScience's GitHub page, you will see a deliberately added statement:
OpenScience is an independent project. It is not affiliated with, endorsed by, or sponsored by Anthropic. “Claude” is a trademark of Anthropic, PBC, used here only to describe compatibility.
To put it plainly: we are an independent project, and we have no connection with Anthropic whatsoever. The mention of "Claude" is purely to describe compatibility, no overthinking needed.
It seems the previous "lobster incident" left a deep impression on the entire open-source community.
After OpenClaw went through several name changes earlier, OpenScience has now solidified this disclaiming statement into the very first version of its README.
The reason is simple: survive first, then talk about being a viable alternative (doge).
Open-source repository link:
https://x.com/i/trending/2073904804829741364?s=20
Reference links:
[1]https://x.com/SynScience/status/2073829478393086311?s=20
[2]https://x.com/i/trending/2073904804829741364?s=20
[3]https://www.openscience.sh/
[4]https://www.anthropic.com/news/claude-science-ai-workbench
This article is from the WeChat public account "QbitAI", written by Yishui, and published with authorization from 36Kr.