Cyberdesk (YC S25), and the accessibility-API alternative for legacy Windows

If you searched this, you probably want two things: confirmation that Cyberdesk is a real Y Combinator company, and a read on how its approach to legacy Windows automation compares to the alternatives. Here is both, verified, with no marketing in the way.

Matthew Diakonov, Written with AI

Published June 22, 20266 min read

Direct answer (verified 22 June 2026)

Yes, Cyberdesk is a Y Combinator company. It is in the Summer 2025 (S25) batch, founded in 2025 by Mahmoud Al-Madi and Alan Duong, based in San Francisco, and as of June 2026 a two-person team. It builds a computer-use agent that lets developers automate legacy Windows apps through an API, and the core is open source. Source: its Y Combinator company page.

The verified profile, as a lookup

Everything below was checked against the public Y Combinator listing and Cyberdesk's own launch posts on 22 June 2026. If you only needed the facts, this is the page in one block.

cyberdesk — public record

What Cyberdesk bets on architecturally

Cyberdesk's model is clean: spin up a full cloud Windows desktop, then control it through a developer API where clicks, typing, and scrolling are function calls. A computer-use agent drives the screen. The clever part of their design is that once the agent has run a task, it replays the memorized steps deterministically and only falls back to a vision model during expected popups, which makes repeat runs faster and cheaper than pure vision on every step.

That is a genuinely good fit for developers who want a programmable sandbox desktop and are comfortable writing the task in code. It is also why Cyberdesk leans on being open source and self-hostable on Kubernetes: teams in healthcare, finance, and insurance need the desktop to live somewhere they control.

Where a different layer fits: reading the accessibility tree

Mediar attacks the same problem, legacy Windows automation, from a different layer. Instead of interpreting pixels on a provisioned VM, it reads the Windows accessibility tree, the same interface screen readers consume. The agent watches a workflow once, then executes it by acting on named controls (a field labeled "Policy Number," a button labeled "Post") rather than on screen coordinates.

That has one practical consequence worth dwelling on. When the UI shifts, a button moves, a window resizes, a label gets reworded, a coordinate or pixel approach has to re-anchor. An accessibility-tree approach keeps working because it never depended on position. There are no brittle pixel matchers and no hard-coded selectors to re-record. You can verify the mechanism yourself in the open-source SDK at github.com/mediar-ai/terminator, which is the same engine that powers the no-code app.

The anchor difference

A coordinate-driven agent answers "click at x=412, y=288." An accessibility-tree agent answers "set the field whose role is edit and whose label is Member ID." The second survives the next UI release; the first usually does not. That is the entire reason the accessibility approach exists.

Side by side

Both are aimed at legacy Windows. The honest framing is that they sit at different layers and suit different buyers. Cyberdesk is built for developers who want a cloud desktop and an API; Mediar is built for ops, finance, and IT teams automating systems that already run on machines they cannot move to a VM.

Feature	Cyberdesk	Mediar
How it reads the screen	Computer-use vision models interpret pixels; clicks and typing are sent as coordinate function calls	Reads the Windows accessibility tree (the same interface screen readers use); no pixel matching, no coordinate selectors
Where the desktop runs	Cloud virtual desktops provisioned on demand (self-hostable on Kubernetes)	Runs on the real Windows machine where the data already lives, on-prem or cloud
Primary user	Developers calling an API and writing the agent task in code	Ops, finance, and IT teams; no-code recorder at app.mediar.ai/web, plus an open SDK for developers who want it
Behavior when the UI changes	Memorized steps replay deterministically; falls back to a vision model on expected popups	Self-heals by matching elements in the accessibility tree by role and label, not by position
Compliance posture	Open source and self-hostable for regulated industries	SOC 2 Type II certified, HIPAA compliant, with audit logs and validation rules

Cyberdesk's facts (batch, founders, open-source repo, architecture) are taken from its public Y Combinator page and launch posts, verified 22 June 2026. Mediar's claims describe its own product.

Which one should you look at?

If you are a developer who wants a programmable cloud desktop, can express your workflow in code, and is happy to host the desktop yourself, Cyberdesk's model is a clean fit and the open-source repo lets you start without a sales call.

If your workflows live on core systems that are licensed and network-locked to specific machines (SAP GUI, Oracle EBS, mainframe terminals, Jack Henry, Fiserv, FIS, Epic, Cerner), and the people who run them are ops and finance rather than engineers, the accessibility-tree approach running on the existing desktop tends to be the lower-friction path. It does not require re-hosting the legacy system, and it self-heals when the UI drifts.

Got a legacy Windows workflow a cloud VM can't reach?

Book a call and we'll show you the accessibility-tree approach on your actual system, not a demo.

Cyberdesk and the alternatives: FAQ

Is Cyberdesk a Y Combinator company?

Yes. Cyberdesk is in the Y Combinator Summer 2025 (S25) batch. It was founded in 2025 by Mahmoud Al-Madi and Alan Duong, is based in San Francisco, and as of June 2026 lists a team of two. Its company profile lives at ycombinator.com/companies/cyberdesk.

What does Cyberdesk actually do?

It gives developers an API to spin up full cloud Windows desktops and control them with clicks, typing, and scrolling exposed as function calls. A computer-use agent runs the task: it replays memorized steps deterministically and falls back to a vision model during expected popups. The stated goal is automating legacy Windows software (EHRs, ERPs) without offshore teams or fragile scripts.

Is Cyberdesk open source?

Yes. The core is published at github.com/cyberdesk-hq/cyberdesk and is self-hostable on any Kubernetes cluster, which is part of why it positions itself for regulated industries like healthcare, finance, and insurance.

How is Mediar different from Cyberdesk?

The architectural bet is different. Cyberdesk drives a cloud virtual desktop with computer-use vision models, sending coordinate-based clicks. Mediar reads the Windows accessibility tree directly, so it acts on named elements rather than pixel positions, and it runs on the machine where the legacy system already lives instead of a provisioned VM. Mediar also targets ops and finance teams with a no-code recorder, while keeping an open SDK (Terminator) for developers.

Why does running on the real desktop matter for legacy systems?

Many core systems (SAP GUI, Oracle EBS, mainframe terminals, Jack Henry and Fiserv banking cores, Epic and Cerner) are installed, licensed, and network-locked to specific machines inside the enterprise. Re-hosting them on a cloud VM is often a non-starter for licensing or security reasons. Reading the accessibility tree on the existing desktop sidesteps that, which is the gap a cloud-desktop API does not cover.

Does the accessibility-tree approach break when an app updates?

Less often than coordinate or pixel approaches. Because Mediar locates controls by their role and label in the accessibility tree rather than by screen position, a button that moves or a window that resizes does not break the automation. There are no brittle pixel matchers or hard-coded selectors to re-record.