OpenScience

Scroll sideways for more →

OpenScience is an open-source, model-agnostic AI research workbench built by Synthetic Sciences, a San Francisco startup in Y Combinator's Winter 2026 batch. It bundles the full computational-research workflow (literature review, hypothesis generation, code and experiment execution, scientific-database queries, and paper writing) into a single locally run application, and it is released under the permissive Apache-2.0 license. ^[1][2][3] Launched in early July 2026, within about a week of Anthropic's Claude Science, the project explicitly bills itself as "a better, open-source Claude Science" and makes model choice, openness, and local control its central pitch against that closed, Claude-only product. ^[1][4]

What it is

OpenScience is described by its makers as an all-in-one workbench that runs the entire research loop rather than just one slice of it. ^[2][7] After a one-line install (npm install -g @synsci/openscience, then openscience), the command-line tool starts a local server and opens an IDE-style workspace in the browser, complete with a file tree, code editor, terminal, and session history. ^[2][8] Inside that workspace an agent reads relevant papers, forms a hypothesis, writes and runs code, executes experiments on real compute, queries major scientific databases, and drafts the write-up, keeping sessions, artifacts, and provenance on disk as shareable links. ^[3][7] The default research agent is supported by domain specialists for biology, physics, and machine learning, plus dedicated critique and literature-review sub-agents and a read-only "plan" mode. ^[4][8]

The tool ships with a large library of what it calls research "skills" (self-contained, editable procedures) and connectors to roughly 30 scientific databases, including UniProt, the Protein Data Bank (PDB), ChEMBL, Ensembl, PubChem, arXiv, OpenAlex, and Semantic Scholar. ^[2][3] Skills span model training (frameworks such as DeepSpeed, PEFT, and TRL), evaluation, dataset work, molecular biology, cheminformatics, LaTeX typesetting, figure generation, and cloud compute. ^[2][10] The workspace renders molecules, protein and genome structures, and plots inline, and it can be extended through the Language Server Protocol, Model Context Protocol (MCP) servers, plugins, and a TypeScript SDK. ^[2][8] As with any autonomous coding agent, this is agentic AI applied to research: a large language model drives a loop of tool calls rather than answering a single prompt.

Origin and Synthetic Sciences

OpenScience is the first public product of Synthetic Sciences, a San Francisco company that press coverage dates to a 2025 founding and that graduated from Y Combinator's Winter 2026 (W26) batch. ^[1][6][10] The launch account posts under the handle "Synthetic Sciences (YC W26)." ^[1] Its guiding claim, repeated across the launch materials, is that "scientific AI should be open" and that "one company shouldn't own the tools the rest of us discover with, or decide who gets to." ^[1]

The company's business model separates the free tool from a paid, optional service. Bring-your-own-key use of the open-source workbench, pointed at the user's own provider accounts, is free and, the company says, never gated; revenue instead comes from Atlas, a managed layer that offers a curated set of hosted frontier models billed from a prepaid wallet, cloud compute, and a persistent shared research graph. ^[1][2][8] The project is careful to note that it is independent and "not affiliated with, endorsed by, or sponsored by Anthropic." ^[2]

Positioning versus Claude Science

The launch was timed and framed as a direct response to Claude Science, the scientific workbench Anthropic released in beta on June 30, 2026 for Claude Pro, Max, Team, and Enterprise subscribers on macOS and Linux. ^[12][13] Claude Science runs only on Anthropic's own Claude models and ships, by Anthropic's count, "over 60 curated skills and connectors" pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics. ^[12][14] OpenScience keeps the same overall shape (an agentic workbench that renders scientific artifacts and traces results back to code) but inverts the trade-offs: any provider instead of one, an open and editable skill library instead of a curated closed one, and local execution on the user's own infrastructure instead of a vendor-hosted app. ^[4][5]

Synthetic Sciences leans hard on that contrast. Its announcement advertises "250+ research skills," "any model ... switching is one flag," and "no throttling, no gatekeeping, no one vendor deciding what science is okay." ^[1] Those comparative and superlative claims, including the implication that 250-plus skills is roughly four times Claude Science's 60-plus, are the company's own marketing rather than independently measured facts, and the two products count "skills" differently. ^[1][4] Reviewers frame the real choice as polish versus openness: Labcritics calls Claude Science "a polished, standalone product with curated integrations" and OpenScience a tool that "trades some polish for openness, auditability, and provider freedom," better suited to computationally fluent researchers and machine-learning engineers who value vendor independence and cost control. ^[4][5]

The face-off sits inside a broader 2026 argument over open versus closed scientific AI, the same debate that surrounds AI for science efforts and OpenAI's science push. A recurring worry OpenScience names is single-vendor exposure: a closed, one-model workbench inherits whatever that model will or will not do, including its provider's safety policies, whereas a model-agnostic tool lets a researcher route around any single vendor's limits. ^[4] The counterpoint, which reviewers also raise, is that Anthropic's curation, hosted compute, and gentler natural-language on-ramp may serve bench scientists who would rather not manage keys, infrastructure, or a developer-style interface. ^[4][5]

Architecture and design

Model-agnostic core. OpenScience routes each request to whichever model the user selects, with no built-in lock-in to a single vendor. Named providers include Anthropic, OpenAI, Google, DeepSeek, GLM, and Kimi; the GitHub README describes support for frontier and open-weight models from "dozens of ... providers," and switching between them is presented as a single flag or selector change rather than a reconfiguration. ^[1][2][8] Because users supply their own API keys, no account is required for self-hosted use. ^[2]

Local and private. The workbench can also drive local models through Ollama, so that no data has to leave the user's machine; the company stresses that OpenScience "runs on your infra" and that "your data stays yours." ^[1][6] This local-first stance is the practical mechanism behind the openness pitch, and it is what distinguishes the workbench from a cloud-hosted product.

Skill packs. The 250-plus skills (the GitHub repository listed more than 290 by mid-July 2026) are plain, readable files organized by domain that users can edit, fork, or extend, in contrast to a closed catalog. ^[1][2] The same openness applies to the agents, connectors, and routing logic, all Apache-2.0 licensed, which is the sort of transparency emphasized by the wider open-source AI movement. ^[2][8]

Atlas and extensibility. Atlas is the optional managed platform (at app.syntheticsciences.ai) that adds hosted frontier models, cloud compute, and a persistent, reproducible research graph shared across many agents; the docs stress that OpenScience "never requires" it and remains fully functional standalone. ^[1][8] Under the hood the project is a Bun and TypeScript monorepo (a Hono/CLI backend with the agent runtime, tools, and skills; a SolidJS browser workspace; a TypeScript SDK generated from OpenAPI contracts; and a plugin runtime), extensible via LSP, MCP, and plugins. ^[2][8] By mid-July 2026 the repository showed roughly 2,400 GitHub stars and more than a dozen releases, reaching v1.3.4 on July 11, 2026. ^[2]

Reception and coverage

OpenScience drew quick pickup from AI-industry press and developer channels. MarkTechPost covered the release on July 5, 2026 (with syndication across several outlets), framing it as an open-source, model-agnostic workbench for machine learning, biology, physics, and chemistry. ^[3] KuCoin's news desk and DailyAIWorld ran explainer and comparison pieces, and Labcritics published a detailed side-by-side against Claude Science. ^[4][5][6] A developer write-up on DEV Community stressed that the tool "runs the whole research loop, not just the reading," and Chinese-language AI directories such as ai-bot.cn and ai-all.info catalogued it for that audience. ^[7][10][11] The project was also posted to Hacker News on July 4, 2026 as "OpenScience: Workbench for scientific research using custom LLMs," where early discussion was modest. ^[9] Much of the coverage echoed the company's own framing, so the openness and skill-count claims traveled largely unaudited.

Limitations and open questions

Several caveats temper the launch enthusiasm. First, maturity: OpenScience shipped only days after Claude Science and is a young, fast-moving project, so most reviews note that any headline demonstrations are early-stage and not yet independently validated, and that both tools help most with the early, computational stages of research rather than replacing human review or wet-lab work. ^[4][5]

Second, verification and self-check. A model-agnostic workbench does not by itself solve the reliability problem, and the reviewer or critique step has a built-in weakness: when the agent that checks the work runs on the same model that produced it, the "reviewer" is not an independent checker. Labcritics makes exactly this point about Claude Science's reviewer agent, and it applies just as much to OpenScience's critique sub-agent, which shares the underlying model unless the user deliberately routes the check to a different one. ^[4] The upside of model-agnostic routing is that a user can point the checker at a second, independent model, but nothing forces them to.

Third, security. OpenScience's agent is not sandboxed; the README states plainly that the permission system "keeps you aware of what the agent is doing" but "is not an isolation boundary," and it recommends running inside a container or VM for sensitive work. ^[2][4] Fourth, usability: the IDE-style, key-it-yourself interface is legible to people comfortable in something like VS Code but presents a steeper learning curve for non-technical researchers than a natural-language app. ^[4][5]

Finally, marketing versus demonstrated capability. The strongest claims ("better," "250+ / roughly 4x," "no throttling or gatekeeping") are the vendor's own; independent, apples-to-apples benchmarks against Claude Science were not available at launch, and skill counts and provider counts vary between the announcement, the README, and third-party pages. Those figures should be read as advertised capabilities, not settled results. ^[1][4]

In plain language (ELI5)

Imagine a robot lab assistant that lives on your own laptop. You tell it a research question, and it reads the important papers, thinks up an idea to test, writes and runs the computer code for the experiment, looks things up in big science databases, and then writes up what it found, all in one window. Anthropic sells a slick version of this called Claude Science, but it only works with Anthropic's own AI and you pay to use it. OpenScience is a free, open version: you can plug in almost any AI brain (including free ones running on your own machine), you can open up and change any of its built-in tricks, and your data never has to leave your computer. The trade-off is that it looks and feels more like a programmer's tool, it is very new, and, like all of these systems, it can still make mistakes, especially when it grades its own homework.

See also

• Claude Science
• Anthropic
• AI for science
• OpenAI for science
• Open-source AI
• Model Context Protocol
• Agentic AI
• Y Combinator

References

1. Synthetic Sciences (YC W26) (@SynScience). "Introducing OpenScience. A better, open-source Claude Science." X (Twitter), July 4, 2026. https://x.com/SynScience/status/2073829478393086311 ↩
2. Synthetic Sciences. "openscience: The open-source AI workbench for scientific research." GitHub repository (README, license, releases), accessed July 2026. https://github.com/synthetic-sciences/openscience ↩
3. MarkTechPost. "Synthetic Sciences Releases OpenScience: An Open-Source, Model-Agnostic AI Workbench for Machine Learning, Biology, Physics, and Chemistry Research." July 5, 2026. https://www.marktechpost.com/2026/07/05/synthetic-sciences-releases-openscience-an-open-source-model-agnostic-ai-workbench-for-machine-learning-biology-physics-and-chemistry-research/ ↩
4. Labcritics. "OpenScience vs. Claude Science: Two Takes on the AI Science Workbench." July 7, 2026. https://labcritics.com/blog/2026/07/07/openscience-vs-claude-science-two-takes-on-the-ai-science-workbench/ ↩
5. DailyAIWorld. "OpenScience vs Claude Science: Open-Source AI Workbench." 2026. https://dailyaiworld.com/blogs/openscience-vs-claude-science-workbench-2026 ↩
6. KuCoin News. "OpenScience Launches as an Open-Source Alternative to Claude Science." July 2026. https://www.kucoin.com/news/flash/openscience-launches-as-open-source-alternative-to-claude-science ↩
7. renolu. "OpenScience runs the whole research loop, not just the reading." DEV Community, July 2026. https://dev.to/renolu/openscience-runs-the-whole-research-loop-not-just-the-reading-3kp ↩
8. DeepWiki. "synthetic-sciences/openscience" (architecture overview). Accessed July 2026. https://deepwiki.com/synthetic-sciences/openscience ↩
9. Hacker News. "OpenScience: Workbench for scientific research using custom LLMs." July 4, 2026. https://news.ycombinator.com/item?id=48786827 ↩
10. AI工具集 (ai-bot.cn). "OpenScience - Synthetic Sciences 开源的 AI 科研工作台." 2026. https://ai-bot.cn/openscience/ ↩
11. AI-all.info. "OpenScience - Synthetic Sciences 开源的 AI 科研工作台." 2026. https://www.ai-all.info/ai-models/openscience-synthetic-sciences-ai ↩
12. Anthropic. "Claude Science, an AI workbench for scientists." June 30, 2026. https://www.anthropic.com/news/claude-science-ai-workbench ↩
13. TechCrunch. "Anthropic's Claude Science bets on workflow, not a new model, to win over scientists." June 30, 2026. https://techcrunch.com/2026/06/30/anthropics-claude-science-bets-on-workflow-not-a-new-model-to-win-over-scientists/ ↩
14. AIwire (HPCwire). "Anthropic Launches Claude Science AI Workbench for Scientific Research." June 30, 2026. https://www.hpcwire.com/aiwire/2026/06/30/anthropic-launches-claude-science-ai-workbench-for-scientific-research/ ↩