Synthetic Sciences has released OpenScience, an open-source AI workbench for scientific research under the Apache 2.0 license. The tool runs on researchers' own infrastructure and is designed as an open, model-agnostic alternative to Anthropic's proprietary Claude Science, which launched in late June 2026.
The team behind OpenScience argues that scientific AI tooling should not be locked to a single vendor. The project is unaffiliated with Anthropic and keeps the full research workflow open, with swappable models and data that never leaves the user's environment.
What OpenScience does
OpenScience is a browser-based workspace backed by a local agent runtime. A researcher gives it a high-level goal. From there, the agent works through a complete loop: it reads relevant literature, forms a hypothesis, writes and runs code, queries scientific databases, runs experiments, and produces a written analysis. The session runs continuously, with all artifacts stored locally.
The workbench is model-agnostic by design. It accepts any frontier or open-weight model - Claude, GPT, Gemini, GLM, Kimi, DeepSeek, and local fine-tunes - routed per request using the researcher's own API keys. No account is needed to start. Installation requires Node.js and one npm command. The first run offers a choice between managed models, your own provider keys, or free demo models.
Four components make the runtime suitable for real work. Research agents include a default agent and specialist agents for biology, physics, and machine learning, plus critique and literature-review sub-agents. 250+ editable skills cover areas such as training with DeepSpeed and PEFT, evaluation, dataset work, cheminformatics, molecular and clinical biology, LaTeX, figure generation, and cloud compute. Scientific databases as tools give agents direct access to UniProt, PDB, ChEMBL, arXiv, and roughly 30 other resources, enabling queries that pull structured data into the workflow. Finally, the workspace itself includes a file tree, editor, terminal, session history, and inline rendering of molecules, structures, genomes, and plots.
Extensibility is built in. OpenScience supports LSP integration, MCP servers, plugins, custom agents, and a TypeScript SDK. An optional managed layer called Atlas provides curated frontier models through a prepaid wallet, a persistent research graph, and cloud compute, but the tool never requires it. All sessions, artifacts, and provenance are stored on disk and can be shared as links. Queries to scientific databases like PubMed or UniProt are central to many Research tasks, and OpenScience treats them as first-class agent tools.
How it compares to Claude Science
Claude Science offers a polished, curated experience with fixed Anthropic models and a set of integrated skills and databases. OpenScience prioritizes openness and provider freedom. Key differences:
- License and vendor: OpenScience is Apache 2.0 open source. Claude Science is proprietary.
- Model choice: OpenScience supports any provider or local fine-tune and lets you switch models per request. Claude Science uses only Anthropic's models.
- Cost model: OpenScience is free with your own API keys and never gated. Claude Science requires a paid subscription.
- Skills and tools: OpenScience ships 250+ skills and 30+ scientific databases as tools, all editable and extensible. Claude Science offers 60+ curated skills and connectors, plus special integrations like NVIDIA BioNeMo.
- Execution environment: OpenScience runs entirely on your infrastructure. Claude Science currently runs on lab machines with beta support for macOS and Linux.
- Sub-agents: Both tools use coordinating agents and specialists, but OpenScience adds a read-only plan mode and more specialist options.
Real-world use
An ML engineer testing a fine-tuning approach can ask the ml agent to pull relevant arXiv papers, use PEFT and TRL skills to write a training script, execute it, and draft a report - all in one session. A computational biologist studying a protein target can query UniProt and PDB, render the structure inline, propose mutations, and track provenance. A chemist screening small molecules can use OpenScience to query ChEMBL and PubChem, filter by bioactivity, and generate ranked candidates with plots. Teams that need to compare model outputs on their own data can route the same task through Claude, GLM, and a local fine-tune without rewriting anything. These workflows illustrate the hands-on AI for Science & Research capabilities the tool brings to a single local environment.
Strengths and limitations
OpenScience's open-source license means every skill and agent is readable and modifiable. Model-agnostic routing eliminates single-vendor lock-in, and local execution keeps private datasets on the researcher's own systems. The 250-plus skills and deep database integrations cover substantial ground across machine learning, biology, and chemistry.
On the other hand, the agent is not sandboxed - the permission system is not an isolation boundary, so researchers who need strong isolation should run it inside a container or VM. The project is young and will have rough edges compared to mature products. Bring-your-own-key means you manage provider costs and rate limits yourself. And the quality of output depends heavily on which model you choose for each request.
Why this matters for science and research professionals
OpenScience gives researchers a practical way to run an end-to-end AI science workflow on their own terms. You choose the models, you keep the data, you control the infrastructure. For labs that handle sensitive data, operate on a tight budget, or need to audit every step of the process, the ability to inspect and modify the tools - and to avoid being locked into a single model provider - is a concrete advantage. The trade-off is that you take on more of the setup and cost management. For those willing to accept that, it offers a level of control that closed-platform alternatives do not.
Your membership also unlocks:
