Minicor Built the API for Software That Was Never Meant to Have One

Somewhere in a hospital right now, a piece of software written before the smartphone existed is refusing to talk to anything modern. No API. No integration. Just a screen, a keyboard, and a human being who copies data into it one field at a time. Minicor, a San Francisco startup in Y Combinator's Spring 2026 batch, built its whole company around that screen.

The pitch is deceptively small. AI companies keep winning enterprise deals in healthcare, automotive, logistics and financial services, then stalling at the last step: their customer's system of record is a legacy desktop application with no way to push data in programmatically. The demo works. The contract is signed. And then the AI has nowhere to put its answer.

Minicor's answer is to make that legacy app behave like an API. You describe a workflow. You get back an endpoint. Every time you call it, a self-healing automation runs on a Windows virtual machine - on-premise, in the cloud, or behind a Citrix login - and types the data in the way a person would, only faster and without getting bored.

01 / The ProblemThe systems no one wants to touch

Enterprise software has a long memory. Electronic health records, dealer management systems, ERPs from the era of beige towers - they run the daily operations of entire industries, and a surprising number of them offer no writable API at all. If you want to get data in, you point, you click, you type.

For a human, that's tedious. For an AI product trying to serve thousands of transactions a day, it's a wall. Companies have tried bridging it with traditional robotic process automation - scripts that mimic clicks - but those scripts are famously fragile. A vendor ships a routine UI update overnight and the automation wakes up broken. Someone has to notice, diagnose, and rewrite it. Multiply that across dozens of customers and, as the founders put it, maintenance becomes the killer.

02 / How It WorksDeterministic code, with a co-pilot that watches the screen

Minicor's technical bet is that neither pure scripting nor pure AI is enough on its own. Hard-coded automations are fast, cheap and reliable until the interface shifts. Pure computer-use agents - models that look at a screen and decide where to click - are adaptable but wobble, landing somewhere around 80-85% click accuracy. In a workflow with dozens of steps, that compounds into failure.

So Minicor runs deterministic Python for the predictable path, and pairs it with AI reflection agents for the moments that go sideways. A reflection agent verifies each action against what's actually on the screen and self-corrects before a mistake cascades. When a vendor changes the UI, the automation adapts instead of crashing. The company reports 93-96% click accuracy in production.

Under the hood, the AI layer leans on Claude and the Model Context Protocol. Minicor ships an MCP server that lets a coding agent navigate a virtual machine, build the automation, and maintain it as the underlying software drifts. It's AI building the plumbing that lets other AI reach the systems it couldn't touch before.

03 / Who Uses ItThe AI company is the customer, not the hospital

Minicor made a deliberate choice about who it sells to. Rather than knocking on the doors of hospitals, dealerships and warehouses directly, it sells to the AI companies that already have those relationships and need a reliable way to write into their customers' systems. That keeps deployments fast and pushes the messy, regulated last mile - compliance included - onto a team that does nothing else.

The practical effect is that a company can go live in days instead of quarters. A logistics startup that promised its customer automatic order entry no longer has to staff a room of contractors to key it in by hand. A dental-software company can offer write-back into the practice management system it never had permission to modify. The AI product keeps its brand and its relationship; Minicor handles the part that used to require a spreadsheet, a temp, and a lot of apologizing.

The list of systems it plugs into reads like a tour of the software that quietly runs the physical economy:

04 / The ModelNothing fails silently, and you pay when it works

Two design decisions define how Minicor earns trust. The first is observability. Every run is recorded on video, logged action by action, and captured with screenshots and execution context. If something breaks, a Slack alert fires. For a hospital deciding whether to let software touch its EHR, that flight-recorder transparency is the price of entry.

The second is pricing. Minicor charges per successful workflow - you pay only when the data actually lands in the target system. A failed automation is free. It's an unusually confident way to price, and it aligns the company's revenue with the one thing the customer actually cares about. It also quietly reframes the sale: the buyer isn't purchasing software licenses or automation seats, they're buying a guaranteed outcome, metered by results.

That structure sits on top of the parts a regulated buyer checks before anything else. Minicor is SOC 2 Type II certified and HIPAA compliant, deployable on-premise so that sensitive data never has to leave the customer's environment. For the AI companies reselling it, those certifications are load-bearing - they're often the difference between a security review that ends in a signature and one that ends in silence.

05 / The FoundersFrom Android Auto and $30M in savings to a Windows VM

Minicor was founded in 2024 - originally under the name Laminar, which still lingers in the company's social handles and its docs.laminar.run domain. CEO Faizaan Chishtie came up as an early engineer at YC startups and once ran a furniture-rental venture. CTO Saheed Akinbile took a longer road: born in Lagos, trained in software engineering at the University of Ottawa, he wrote Java at Morgan Stanley that reportedly saved over $30 million a year in fees, then spent three years at Ford building Android Auto drive modes before turning to legacy-system integration full time.

It's a pairing suited to the problem. The work isn't glamorous - it lives in the depths of software most engineers would rather never open - but it's exactly the kind of unfashionable plumbing that becomes valuable the moment everyone else is racing toward the shiny layer on top.

06 / The LandscapeWhere it sits

The obvious comparison is to traditional RPA vendors like UiPath and Automation Anywhere. Minicor's argument against them is the same one it makes against in-house automation teams: script-based approaches are brittle and expensive to keep alive. Its argument against pure computer-use agents is the accuracy gap. The claim is that the hybrid - deterministic where it can be, adaptive where it must be - is the only version that survives contact with real production volume.

Whether that holds as the company scales beyond healthcare is the open question. But the underlying observation is hard to argue with. The frontier models get the headlines; the systems of record don't move. Someone has to stand in between them. Minicor decided to be that someone.