EN
Back to the archive

The archive · Developer & Business Tools · Product decision · 2026

Minicor bets deterministic plus recovery agents beats pure AI for legacy Windows work

YC P26 startup Minicor sells deterministic desktop automation with AI recovery so AI products can enter legacy Windows systems; HN launch, May 2026

Minicor

The betThat AI enters legacy industries not via a computer-use model improvising every click, but via deterministic desktop automation with agents for recovery and edge casesLive

What the business is

Minicor is a YC-backed platform that lets AI products read and write legacy Windows desktop systems — EMRs, dental PMS, automotive DMS, SAP — through one API call, with no customer APIs needed.

How it started

Minicor is a YC Spring 2026 batch company that launched publicly on Hacker News on May 26, 2026. Its founding observation was that most systems of record in healthcare, logistics, automotive, and finance are old Windows desktop apps without APIs, so AI products selling into those industries live or die on integration.

What happened

The platform runs workflows as deterministic code rather than letting a model improvise every run, adding a reflection agent that verifies each click against what is on screen and recovers from UI changes or surprise dialogs. Founder comments on the HN thread explain the hybrid rationale: pure computer-use models are slow, token-hungry, and risky for critical tasks like updating patient records, while manually maintained RPA code breaks constantly — so AI writes the code and monitors it instead.

How it ended up

As of its May 2026 launch, Minicor is live with customers running workflows in production and is positioning for regulated rollouts with SOC 2 Type II, HIPAA, Citrix, and on-premises deployment support. The company is still early, without named customers or disclosed revenue.

Background

Minicor, a YC Spring 2026 startup, launched on Hacker News on May 26, 2026 with a blunt thesis: most systems of record in healthcare, logistics, automotive, and finance are legacy Windows desktop applications with no APIs, and AI products selling into those industries die on integration.

Its platform turns those desktops into something programmable: one API call triggers a full workflow on a Windows VM, with the workflow stored as deterministic code and a reflection agent verifying each click against what is on screen and recovering from UI changes. Minicor argues that pure computer-use models are too slow and risky for critical tasks like updating patient records, while traditional script-based RPA breaks constantly when maintained by hand — so it uses AI to write and monitor durable code instead.

The company claims 93–96% click accuracy in production versus 80–85% for other approaches and says the architecture already runs at 25,000 patients/day, though it names no customers. The bet is that reliable desktop automation is the shortest path to revenue for AI vendors selling into legacy-heavy sectors, and that compliance artifacts like SOC 2 and HIPAA will matter more than demo magic.

What has to be true

  • Legacy systems of record in healthcare and industry have no APIs, so AI companies need someone to make the old Windows apps programmable before they can ship
  • Pure computer-use agents are slow, token-hungry, and risky for critical writes, while deterministic code fails when UIs change — Minicor combines code with AI recovery
  • Selling to AI vendors rather than hospitals and factories lets Minicor ride deals its customers already won instead of running its own enterprise sales cycle
  • The 25,000 patients/day production claim and 93–96% accuracy figures, even unaudited, show the wedge is working in real regulated workflows

What can be applied

Making the old system programmable often beats replacing it: deterministic code plus AI recovery agents turns brittle RPA into a reliable way for AI to enter legacy industries.

Aftermath

As of the May 26, 2026 launch, Minicor is live and taking customers: the platform runs on Windows VMs or in browsers, on-premises, in the cloud, or inside Citrix environments, and the company reports production use at 25,000 patients/day. Founder activity on the HN thread indicates active iteration on trust-center and documentation feedback, with the next phase aimed at AI scribes and other vendors integrating into hospital EMRs, automotive DMS, and financial systems. Revenue and customer names have not been disclosed.

Sources

spotted an error? The archive wants to know.

Your turn

You just read one. Describe what you are building, and see who is betting on the same thing.

Free account · 3 free questions · no card

Related cases