The benchmark number jumped out first: 82%. That's what Coasty, a Y Combinator startup with two founders and zero venture capital, says its computer-use agent scored on OSWorld, a benchmark designed to test whether AI systems can actually navigate real desktops—clicking through Ubuntu interfaces, filling out forms, managing files—the way a human operator would.
If accurate, that figure would put the fledgling company ahead of OpenAI's Operator and Anthropic's much-publicized computer-use features, at least on this particular test. OpenAI's tool, according to comparison data Coasty has compiled, landed somewhere in the high 30s to low 40s on the same benchmark—though OpenAI hasn't published official OSWorld numbers. Anthropic's system reportedly scored in the 60–70% range across similar evaluations, though those figures come from Coasty's own marketing materials rather than official disclosures.
It's a bold claim for a company that, as of a few months back, described itself on LinkedIn as "Two founders (Stanford x Columbia). No funding." Whether that claim holds up under scrutiny—and whether OSWorld scores translate to real-world reliability—is another matter entirely.
Vision Over Selectors
Coasty's pitch centers on a familiar pain point in enterprise automation. Traditional robotic process automation tools—think UiPath, Automation Anywhere, the legacy platforms that helped companies digitize repetitive tasks—rely on what developers call selectors. These are specific identifiers tied to buttons, text fields, and other interface elements. Change the layout of a web form, update a software UI, and the automation breaks. Someone has to go back in, reconfigure the script, test it again.
Coasty's approach sidesteps that fragility, at least in theory. The agent interprets screen pixels directly, the way a person would. It watches the desktop, predicts the next action—click here, type this, scroll down—and executes. No DOM selectors. No element IDs. Just vision.
The system operates on secure cloud virtual machines or locally through a desktop app, depending on the use case. Developers send base64-encoded screenshots and instructions to an API; the model returns actions. Sessions hold state across multi-step workflows. The company supports parallel agents, scheduled routines, webhook callbacks—features designed for production environments, not just demos.
It sounds straightforward. But the devil, as always, lives in the execution.
The Benchmark Question
OSWorld, developed in 2024, tests agents on real Ubuntu desktop tasks—multi-application workflows, file management, terminal commands. It's become something of a proving ground for companies building computer-use agents, though it's worth noting that benchmarks don't always reflect production complexity. Controlled test environments rarely do.
Coasty's marketing materials cite that 82% score repeatedly, a figure mentioned in a Heavybit article from June. Yet a LinkedIn post from around May, announcing the company's v3 model, claimed 84%. The discrepancy is small, maybe just a rounding difference or an updated evaluation, but it's there.
More notable: no independent, third-party leaderboard surfaced in research that verifies Coasty's exact performance. Benchmark aggregators like Epoch and BenchGecko track OSWorld generally, but they don't catalog proprietary agent frameworks like Coasty's. The 82% figure appears to come from the company's own testing, reported by Heavybit but lacking independent verification.
And then there's OSWorld 2.0. Released on June 28, the updated benchmark introduces longer-horizon tasks—sequences exceeding 250 steps, designed to push agents further. No Coasty score on that version has emerged yet.
Two Models, Different Approaches
Under the hood, Coasty runs two distinct models. Version 3 operates in a guided, step-by-step mode. Version 4 takes a more autonomous route, with a built-in verifier that checks its own work as it goes.
Developers provision Linux or Windows virtual machines through the API. Pricing breaks down to roughly five cents per hour for Linux VMs, nine cents for Windows—calculated from Coasty's rate of one cent per credit, with credits consumed at different rates depending on the OS.
The system integrates with development tools: VS Code, Cursor, Claude Desktop. JSON workflows define task sequences. Session memory persists across steps. There's even an "on_awaiting_human" control, which lets developers set handoff points where the agent pauses and waits for a person to intervene.
The company open-sourced part of the stack on GitHub, branding it as the "State of the Art 82% OSWorld Computer Using Agent." Whether that component lives up to the tagline depends, again, on how teams deploy it.
Free to Start, Metered at Scale
Coasty offers a free tier—no credit card required. Paid plans scale based on concurrent agent sessions and VM hours. Management endpoints and event streaming come at no charge; metering applies only to compute resources.
In March, the company introduced Coasty Desktop, a local version that lets the agent control machines directly. That matters for use cases involving files or applications that can't easily run in a cloud environment—legacy desktop software, local databases, anything tied to specific hardware.
The pricing model suggests Coasty is angling for production deployments, not just experimentation. Whether customers are actually committing remains unclear. The company hasn't disclosed named enterprise clients or announced a seed round.
A Crowded, Fast-Moving Space
Coasty isn't operating in a vacuum. OpenAI launched Operator in January 2025. Anthropic rolled out computer-use capabilities around the same time. Google's Project Mariner made headlines before being quietly discontinued. Browserbase, another infrastructure player, announced its own agent platform on June 30, just days before this story.
Meanwhile, the RPA incumbents aren't sitting still. UiPath and others have started layering what they call "agentic AI" atop existing automation frameworks—an acknowledgment, perhaps, that the market is shifting toward more adaptive systems.
Coasty's blog posts from March through June make the case directly. Traditional RPA breaks when interfaces change. Vision-based agents adapt. Vertical-focused pitches target use cases in reporting, marketing operations, e-commerce fulfillment, HR recruiting—anywhere repetitive tasks pile up.
It's a compelling argument. But RPA vendors have decades of enterprise relationships, compliance certifications, and support infrastructure. Displacing them won't happen quickly, no matter how impressive the benchmarks look.
The Team Behind the Tool
Prateek Jannu and Nitish Kovuru, the two founders, both list Stanford and Columbia credentials on LinkedIn. Jannu's profile notes he turned down a Stanford admit to build Coasty. Kovuru mentions TA experience for a graduate AI course and prior work on enterprise sales automation.
The YC directory places them in the Summer 2026 batch, though LinkedIn shows the company founded in 2024—possibly a gap between incorporation and acceptance into the accelerator.
Product Hunt metrics from the March 12 launch show 84 upvotes and a #44 ranking that day. For a developer tool, those numbers suggest modest early traction, not a viral moment.
No announced funding. No disclosed customers. Just two people and a product that, if the benchmark holds, performs better than what OpenAI and Anthropic have shipped to date.
What Comes Next
Whether 82% on OSWorld means anything in production is the real test. Benchmarks capture a snapshot under controlled conditions. Real automation workflows involve edge cases, unexpected UI changes, flaky network connections, and all the messy variability that enterprises deal with every day.
Coasty's API is live. The documentation is thorough. The benchmark claim, assuming it's accurate, is striking. But automation has always been as much about reliability and support as raw capability. A system that works 82% of the time in a test environment might still fail in ways that matter when real money is on the line.
For now, Coasty exists in that uncertain space most early-stage infrastructure startups occupy: promising enough to attract attention, unproven enough that adoption remains a question mark. The founders are betting that vision-based agents represent the next phase of automation, that the brittleness of selector-based RPA has become untenable.
Maybe they're right. Or maybe, like so many ambitious technical pitches before, the hard part turns out to be not building the demo but surviving contact with the market.
