“8pm. 30 patients. 50 jobs. One crumpled sheet of paper.” That is how Khalid Shamiyah, an obstetrics and gynaecology resident in Oxford, explains why he built Handover, a free iPhone app for ward jobs that lives entirely on the phone and hands your list to a colleague by QR code at the end of the shift. It was one of 46 products featured in The Handover between 7 August and 2 October.
I wanted to know what those 46 have in common, so I went further. Over the same nine weeks the digest behind this site caught another 62 concrete artefacts built by clinicians: apps, libraries, simulators, analysis packages, practice systems. I classified all 108 by hand on 32 columns: what it is, who built it, where they are, how it is deployed, whether it touches patient data, whether AI is inside it, what evidence exists, what the builder says about safety. This piece is what the census shows. The data lives in a table now, so we can repeat it in January and see what moved.
One warning before the numbers. This is a count of what surfaced in English on X, GitHub, Reddit and a few journals, filtered through one editor. It under-samples Japan, China, Latin America and francophone Africa, and almost everything in it is self-reported. Treat it as a field census, not a market study.
The shape of it
Three clusters hold two thirds of the set. Teaching tools, which means simulators, games and courses: 23. Plumbing, which means developer tools, terminology and standards toolchains and reproducible research code: 26. Running the shop, which means practice management, telehealth and ward or admin workflow: 18. Patient-facing trackers are 9, decision support is 9, and a third of all 108 have no disease or specialty in them at all.
Geography: 45 builders in North America (43 in the United States), 30 in Europe (12 in the UK, 5 in France, 3 in Italy), 14 in Asia, 4 in Africa, 3 each in South America and Oceania. 78 physicians, 15 nurses, 9 pharmacists, 3 dentists, 2 physician assistants. 97 of the 108 built alone.
They build for their own shift, not for patients
55 of the 108 artefacts are for clinicians in the builder's own role. 26 are for students and trainees. 20 are patient-facing. 41 builders say outright that the origin was their own workflow pain. Zachary Meade, an anaesthesiology resident, built MMEasy, an opioid conversion calculator, because he was doing the maths on scrap tape. A CVICU nurse spent the better part of a year on Critical Care Vault, a bedside reference he wanted on his own unit. Karthik Deegutla, an anaesthetist in India, built a free case logbook in about eight hours because the paid ones annoyed him.
The product the public conversation keeps imagining, an AI doctor for patients, is almost absent. What clinicians build is the thing that was missing at 8pm.
The biggest cluster is teaching, not treating
23 of the 108 are simulators, games or courses. An ECG synthesiser with physiological modelling. ED Rush, an emergency-department simulator with over 130 cases, where every patient's board timer turns amber and then red. An aneurysm-clipping simulator a neurosurgeon built in 24 hours. Oral-boards coaches for urology and emergency medicine. A browser-based cardiac-arrest trainer by a nurse in Kuwait. Open Vent Sim, a ventilation simulator Italian and Czech anaesthetists built in about 40 hours and then validated in residency training.
Clinicians know exactly where their own training was thin, and they now build the missing exercise. This is also the only cluster where validation is actually happening: the two artefacts "used in formal training" and most of the peer-reviewed ones sit here.
A quarter of it is invisible plumbing
26 artefacts are libraries, SDKs, terminology toolchains and analysis packages. Josh Mandel's mychart-takeout exports your own health data from an Epic patient portal without a language model anywhere in the path. Marcus Baw's sct and Mark Wardle's hermes are two SNOMED CT toolchains, both from the UK. Dan Heslinga keeps shipping Kotlin FHIR libraries for the WHO Open Health Stack. Blaine Warkentine's SolvingHealth SDK is "free pipes for healthcare AI".
These are not vibe coders discovering code. They are the clinician-developer old guard, people who were already professional-grade, now shipping faster. The census only makes sense if you keep the two populations apart: the 27 builders with formal or professional coding backgrounds and the 11 who say they had never written a line of code before this year. Praneet Mylavarapu's line is the bridge between them: the prototype becomes the spec.
AI built it, but AI is mostly not in it
This surprised me most. 59 of the 108 products contain no AI at all. 28 run a large language model at runtime, 6 run a classical model, 2 use AI only in an offline content pipeline. The vibe-coding wave is producing mostly deterministic software: calculators, charting, handover lists, simulators, libraries.
Several builders say so on purpose. Critical Care Vault states there is no language model inside the shipping app. mychart-takeout replays the portal's own API calls and nothing else. Bartosz Fiałek's RheumaNext "deliberately runs no LLM at runtime". Sohrab Arora's oral-boards coach grades on a deterministic rubric because "this app is not just a wrapper around a chatbot". Stefano Martano's prescription-refill flow uses a deterministic check before a doctor clicks accept. A specific kind of clinical judgement is showing up in architecture: where hallucination would hurt, these builders designed the model out.
Of the 28 products that do run a language model at runtime, 22 have no validation of any kind and 7 handle identifiable patient data.
Twenty products hold real patient data, nineteen say nothing about assurance
20 of the 108 handle identifiable patient data: practice-management systems, a telehealth pharmacy in Australia built by a pharmacist who cannot write a line of code, a GP platform in Yorkshire built for £175, ICU charting in Russia, and a 48,000-line clinical system a Japanese dentist started building in July that has run hospital-wide for a month.
Nineteen of the twenty say nothing about regulation or clinical safety assurance. The exception is ketviewer, Keith Grimes's local viewer for NHS out-of-hours messages, published explicitly as a proof of concept without DCB0129 or DCB0160 assurance. Open Vent Sim does the same from the other side: "requires professional technical oversight for production use". Both British.
The gap is specific. 21 of the 108 builders volunteer a safety claim, and "data stays on the device" is the most common one, followed by "no account" and "no LLM at runtime". Awareness of safety in general is there. What is missing is awareness of clinical safety assurance in particular: the amber and red lanes of the founding essay, the DTAC, the DCB standards, Article 5(5). Nobody is hiding anything; the vocabulary simply has not reached them. That is the job of this site's Build safely section and of guides like which platforms sign a BAA.
Europe keeps it local, America ships SaaS
Of the 20 identifiable-data products, 8 are European and 8 are North American. Four of the European eight are local-first or self-hosted: DentVault keeps dental records on the machine, MediFlow runs fully offline for an Italian district doctor, ketviewer never leaves the practice, RemCard is source-available and self-hosted. Seven of the eight American ones are cloud-hosted, and three of those are positioning as platforms for other doctors: WithCline, a "clinical harness" for HIPAA-compliant apps; KaufCare, a full SaaS suite; TJM Labs' pharmacy agents.
Europe is also where the standards tooling comes from: the two SNOMED toolchains and an EHDS patient-sovereignty proxy that implements Articles 3, 7 and 8 of the European Health Data Space in FHIR consent masking. A transatlantic difference in instinct, visible even at this sample size, and it maps onto GDPR and NHS culture on one side and the HIPAA-compliant SaaS market on the other.
Solo, free and unvalidated is the modal product
97 of 108 built alone. 62 free, 13 with any paid model at all. 77 with no validation. 44 are live with none. Money is not the motive, peer review is rare (6), and App Store ratings are the most common proof of life. "Clinicians are the new builders" is true as a statement about who is making things. It is not yet true as a statement about who is making things that have been checked.
Career stage predicts the product better than credential does. Practice owners (12) and clinicians who have left practice (11) make the business-shaped things: telehealth, platforms, education businesses. Residents and students (19) make the research packages, coursework and ward tools. Attendings (36) make the simulators and references. Nurses, pharmacists and dentists, 27 of the 108 together, build closer to the bedside and the till than physicians do: drug sheets for French nurses (Bolus), a TB-adherence game for children in Eswatini (PhiliGo), practice software for a concierge pharmacy.
What nobody shipped
In nine weeks, no clinician in this set shipped a product with a regulatory claim, a device classification or an external clinical evaluation. Disease-specific tools are a small minority: TB adherence, four skin conditions, insulin titration, opioid conversion, stroke triage. Nobody claims a live EHR integration; the one SDK that could says plainly that "no production EHR integration is live". The absences draw the frontier for the next six months more sharply than the presences do.
The census continues
Everything above now sits in a table behind this site, linked to the directory and to every Handover issue, so the January count will be a query rather than a hand pass. If you built something in this window and it is not in the 108, tell me and I will add it. If you are in the twenty and want to know what the amber lane actually asks of you, that is what the guides are for. And if you are still deciding whether to build: the modal clinician builder in this census is one person, working alone, fixing the thing that annoyed them at 8pm. You qualify.