MatOS is an agent operating system tailored for materials research, addressing the challenges of scattered data, heterogeneous tools, and multi‑day workflows that span literature analysis, computation, and experiment. The system is built around three core components:
- Layered scientific asset graph – data, knowledge, tools, skills, roles and workflows are abstracted as reusable assets and linked through a dependency graph.
- Stateful adaptive workflow – tasks run as explicit execution graphs equipped with scientific checkpoints, allowing interruption, resumption and full traceability.
- Evidence‑driven asset evolution – execution logs and user feedback automatically refine assets, enabling continual learning and capability growth.
On the ResearchClawBench benchmark, MatOS improves the previously best‑performing agent by 11.15% while keeping the same base model, and expands the cost‑score Pareto frontier. With the same base model it outperforms Codex CLI by 2.13% on MatTools and by 4.17% on SciAgentGym. In eight real research cases, MatOS narrowed more than 30 million candidate structures to 97 solid‑electrolyte candidates, identified 89 halide variants with higher computed conductivity, linked force errors of 11 pretrained interatomic potentials to local atomic environments, and supported two new methods that reduced errors by 3.1% and 13.6% respectively.
The code is open‑source at https://github.com/DingyangLyu/MatOS
Review