Launch HN: Coasty (YC S26) – An API for computer-use agents

Nitish and Prateek's pitch for Coasty starts from a familiar complaint: half the software running big institutions has no usable API. Payer portals in healthcare, ancient Windows accounting tools, internal systems that only talk to a mouse and keyboard. Traditional RPA tried to solve this by recording clicks and replaying them, and that works right up until a button moves or a popup appears, at which point the whole script falls over.

Coasty's answer is to skip scripted replay entirely and let a vision-based agent look at the screen, decide on an action, execute it, then look again before deciding the next move. No DOM access, no accessibility tree, no per-app integration. That's the same trick that lets it operate a browser, a remote desktop, or a legacy Windows app through one API, since it never actually reads the underlying code, just the pixels.

Developers send a task in plain English — open this record, enter this data, don't submit if the ID doesn't match, return the confirmation number — along with files and credentials, and Coasty runs it inside an isolated VM. What comes back isn't just a pass/fail. It's a structured run record: final status, extracted outputs like a confirmation number, a replay URL, and a timestamped log of every verification step the agent took. Developers can also force a pause for human approval before anything irreversible happens, like a final submission.

The founders say the real engineering problem isn't getting an agent to complete a workflow once — that part is almost easy now that vision models are decent. It's getting it to do that reliably, over and over, without quietly typing the wrong patient name into a field because a page loaded slow. So they've built a layer that tracks expected state, flags when the app has drifted from it, and lets developers define invariants like