albertcheng19/medskillos

MedSkillOS is an open-source framework for adding structured medical and biomedical knowledge to AI agents. Instead of giving an AI agent a collection of loosely organized prompts, this project provides a standards layer: each medical skill comes with defined input and output formats, safety boundaries that limit what the agent should and should not attempt, and records of what evidence was used and what the agent was uncertain about. The stated goal is to make AI agents more useful and more reviewable when working in medical or research contexts. The system is built around four layers. The first is a set of domain skill packs, which group related skills by medical area. The initial focus is on clinical diagnostics (such as building a differential diagnosis list, flagging warning signs, and summarizing findings for different audiences like doctors, nurses, and patients) and on clinical neuroscience (processing brain recording data from EEG, MEG, and fMRI instruments). Planned areas include literature review, study design, bioinformatics, scientific writing, and drug safety. The second layer is a harness that runs each skill and checks that inputs and outputs follow the expected formats, that safety boundaries are respected, and that the results can be reproduced and inspected. The third layer is a set of structured objects for recording evidence, reasoning steps, and reviewer decisions, so that results are not just free text. The fourth layer supports controlled skill improvement: proposed changes must pass scope, safety, and regression checks, and be reviewed before they move through a defined maturity chain from draft to stable. The project describes itself as being in early development. It is MIT-licensed and open to contributions.