The YC S26 startup ships an API that rents developers a managed machine where an AI can move a mouse and drive real apps the way a person would.
Computer-use agents are a new class of AI software that drives a computer the way a person does: moving a mouse, opening apps, filling forms, and reading whatever is on the screen. The current generation of agent tooling usually meets the agent halfway: a browser tab, a headless DOM, or a parsed accessibility tree. Coasty, a Y Combinator S26 launch that hit Hacker News this week, is making a different bet. The startup rents developers a managed Windows or Linux machine and exposes a single HTTPS API for sending a goal, watching the agent act, and stepping down into lower-level controls when needed.
The pitch, on Coasty's Hacker News launch thread and the company's own product page, is that a screen-actuated agent can reach the long tail of software a browser session cannot. If the agent sees pixels and acts on coordinates, the same API covers a SaaS browser tab, a remote Windows desktop, and a legacy admin tool nobody has touched since 2014. There is no DOM, accessibility tree, or selector strategy to maintain. The cost is a heavier unit of account: a full virtual machine per run, billed while the agent works.
The API surface, per Coasty's docs, is intentionally narrow. Developers send a goal to POST /v1/runs with an idempotency key, and Coasty returns a terminal state. Above that sits a workflows layer for chaining runs. Below it sit lower-level primitives: sessions, predict, grounding, and parse, for cases where the developer wants to drive the agent by hand. Authentication is X-API-Key or Authorization: Bearer, and the public test keys (sk-coasty-test-*) hit mock virtual machines and never bill. The shape is closer to a hosted model API than to a browser-automation library.
The benchmark Coasty points to is OSWorld, a public evaluation from HKU, Salesforce, CMU, and Waterloo that runs an agent through 369 real desktop tasks across Chrome, LibreOffice, GIMP, and other apps. The benchmark is real and active: the project's site lists OSWorld-Verified as the live version and shipped OSWorld 2.0 on 2026-06-26. Coasty's marketing page claims its in-house model ranks first on OSWorld at 85.6%, with 82.81% independently verified on the official leaderboard. Both numbers are Coasty's claims, not this article's, and they should be re-checked against the current OSWorld-Verified leaderboard before any downstream reporting treats them as fact.
On the commercial side, Coasty's customer-support page makes the company's own case. It cites a 76% auto-resolution rate for a Coasty-powered support workflow, alongside industry figures the page attributes elsewhere: roughly 80% of support tickets are routine, a human-handled ticket costs $20 to $30, support turnover runs 30 to 45% a year, and more than 65% of chatbot sessions end without a resolution. The 76% number is a company claim, not an independent benchmark, and the surrounding statistics are co-marketing framing rather than peer-reviewed evidence. The HN post also includes a demo video the founders flag as a mockup, not a production run.
The honest counterargument is structural. A managed-desktop runtime is heavier than a browser-automation API, and the agent ecosystem has so far tilted toward browser-scoped tasks. If buyers converge on web workflows, the desktop lane may stay niche. Reliability is a separate question: the API docs document an approval-required guardrail pattern for high-stakes actions such as the prior-authorization entry the founders demo into a payer portal, but the platform is one early entrant in a crowded lane that already includes browser-automation stacks and in-house agent frameworks. Pricing at scale is undisclosed.
Coasty is shipping public API keys and a docs site this week, with the HN launch as the primary signal. The watch items are concrete: a third-party leaderboard entry on OSWorld-Verified, a named paying customer beyond the founders' own demos, and a public pricing page. Until those land, Coasty is best read as one early bid to host the agent runtime on a remote computer, not the runtime itself.