The question
Can clinicians spend less time hunting, reconciling and re-entering information without losing provenance, uncertainty or human control?
Clinical AI · interoperability · safety
Return time to care — without weakening trust.
CareOS explores a layer above fragmented hospital systems: source-linked clinical context for clinicians and bounded AI agents, with provenance, uncertainty and human authority kept explicit.
Try the clinician demo →
Explore the healthcare endgame →
See the Recare work sample →
Inspect the engineering →
Synthetic / pre-hospital research only. No identifiable patient data. No production write-back.
Three ways in.
The project is deliberately layered. A clinician should not need to read the architecture. An engineer should be able to go all the way down to policy boundaries, fixtures and regression tests.
See the workflow.
A synthetic Infectiology case focused on microbiology lifecycle, documented therapy, pending work, contradictions and source inspection.
Open clinician demo →Attack the agent.
Wrong patient, prompt injection, source outage, stale results and unauthorised writes become replayable safety cases.
Open Recare work sample →Inspect the contracts.
FHIR/ISiK, provenance, clinical state, deterministic authority, hospital rollout and global portability live in the repo.
Open repository →Connect. Don’t replace.
CareOS keeps existing systems of record authoritative. The research question is what trustworthy context and agent controls should exist above them.
KIS · LIS · FHIR · documents · ePA
↓
source-linked context
identity · provenance · time · state
contradiction · freshness · pending work
↓
deterministic policy boundary
↙ ↘
clinician UX bounded agent
↘ ↙
human decisionEvidence before claims.
No “9.8/10 ready” badges. Each layer is labelled by what has actually been demonstrated and what still requires a hospital, vendor or clinician.
Open collaboration · Apache-2.0
Healthcare is a team sport.
Software engineers, clinicians, designers, security researchers, interoperability specialists and researchers are invited to contribute. Start with one bounded problem: one failure → one fix → one regression test.