openscience

OpenScience is a research workbench, not a chatbot with a science theme. The loop it demonstrates on its own homepage is the product: you ask which T4 lysozyme point mutants are predicted to be stabilizing, and the agent loads a research-lookup skill, searches two sources, reads a file, runs `python ddg_scan.py --pdb structures/2LZM.pdb`, and returns a scatter plot with r = 0.71 against 26 ProTherm measurements — plus the three stabilizing substitutions, their predicted and measured ΔΔG values, and an offer to run FoldX next or draft the methods paragraph. Thought, Working and Completed steps are visible, and any run can be copied, undone or forked, which is what makes an agent's output auditable enough to put in a paper. The breadth is real. Databases are exposed as first-class tools rather than a search box: UniProt, PDB, ChEMBL, PubChem, arXiv and 37 more. On top sit 371 skills across biology, chemistry, physics, ML and writing, with a curated research core, and the docs describe adding your own research procedures. Compute is managed — it builds environments and scales from a laptop to a cluster or GPUs — so a scan you prototyped locally doesn't need rebuilding to run bigger. Where it gets complicated is model access. There are four paths: free included models, your own provider API keys, an OpenAI sign-in that reuses an existing ChatGPT Plus/Pro subscription, or a local model via Ollama, LM Studio or another compatible endpoint. You can switch model access without starting a new project. Sitting alongside this is Ace, a handpicked set of models OpenScience has benchmarked for scientific agents with managed search and memory behind one Wallet — a separate commercial layer whose whole pitch is that routed performance is otherwise inconsistent. That's a fair concern stated plainly, but it's also an upsell, and your free-model results won't match Ace's. Interface choice is genuine: a desktop app for macOS, Windows or Linux, a browser workspace after installing the CLI, and a terminal single-turn mode (`openscience run "your request"`), plus an SDK and local API, MCP, and custom tools. Install is one line via curl or npm. The weaknesses are the ones the docs admit. Someone on your team owns setup and maintenance — environments, compute, keys — so a lab with no technical owner will struggle. The benchmark table is vendor-run, and the most-quoted number, 47.3% pass@3 on OpenScience Bench, is the least comparable because the vendor defines the test. Ascent appears in the navigation but is not described anywhere on the page, so don't plan around it. And the vendor is InkVell Inc. (dba Synthetic Sciences), which matters if procurement needs a paper trail. Where it fits: computational groups who already hold API keys or a ChatGPT Plus/Pro seat, labs with data-governance rules that can self-host, and PIs standardising model and skill configuration across a team. Where it doesn't: anyone who wants to type a question and get an answer with no