Computer-Use 2.0 Says the Screen Was Never the Point
trycua/cua jumped to number two on GitHub trending today, 1,124 stars in a day on top of 24,318. The pitch on the repo is four words that most of this industry has been avoiding: computer-use 2.0. Their definition is an agent that moves between code, APIs and graphical interfaces inside the same task, instead of picking one and pretending the other two do not exist.
The piece that makes it interesting is CUA-S1, which also showed up as a Show HN today. It is a family of small specialized models for what the team calls System 1 decisions. Fast, bounded, low-stakes: what value goes in this form field, is this element the thing I should click. The model does not write the answer token by token. It scores choices from the structured interface elements it is handed. The first research profile they shipped is forms, which is about as unglamorous as computer use gets and also about as common.
Read that against what everyone else has been building and the bet is clear. Two years of work went into teaching enormous general models to look at pixels and click correctly. Cua is saying most of those clicks were never a reasoning problem. They are a classification problem wearing a reasoning problem's clothes, they cost frontier-model prices, and they hallucinate because a text generator can always produce a field value that does not exist. Fix the output space and that failure mode is gone by construction.
The rest of the stack is the unsexy part that makes the bet testable. Cua Fleets gives you cloud-isolated desktops through run.cua.ai. Cua Driver does cross-platform app automation from a CLI, MCP or an SDK, on macOS, Windows and Linux. Lume runs local VMs on Apple Silicon. Cua Bench builds tasks, evaluates agents, exports trajectories. That last one matters most, because a company shipping small specialized models needs the training data, and the harness is where the training data comes from.
What makes this worth writing down is that it is the second team in a week to land on the same idea from a completely different direction. TypeSafe AI shipped Jev two days ago, a transformer that outputs calibrated probabilities instead of text, on the argument that human language is the wrong interface for automation. Cua arrived at the same place from the desktop side. When two teams who are not talking to each other both decide the big model should be routed around for the boring 90 percent, that is not a product opinion anymore. That is the shape of next year's cost structure. Repo at github.com/trycua/cua.
← Back to all articles
The piece that makes it interesting is CUA-S1, which also showed up as a Show HN today. It is a family of small specialized models for what the team calls System 1 decisions. Fast, bounded, low-stakes: what value goes in this form field, is this element the thing I should click. The model does not write the answer token by token. It scores choices from the structured interface elements it is handed. The first research profile they shipped is forms, which is about as unglamorous as computer use gets and also about as common.
Read that against what everyone else has been building and the bet is clear. Two years of work went into teaching enormous general models to look at pixels and click correctly. Cua is saying most of those clicks were never a reasoning problem. They are a classification problem wearing a reasoning problem's clothes, they cost frontier-model prices, and they hallucinate because a text generator can always produce a field value that does not exist. Fix the output space and that failure mode is gone by construction.
The rest of the stack is the unsexy part that makes the bet testable. Cua Fleets gives you cloud-isolated desktops through run.cua.ai. Cua Driver does cross-platform app automation from a CLI, MCP or an SDK, on macOS, Windows and Linux. Lume runs local VMs on Apple Silicon. Cua Bench builds tasks, evaluates agents, exports trajectories. That last one matters most, because a company shipping small specialized models needs the training data, and the harness is where the training data comes from.
What makes this worth writing down is that it is the second team in a week to land on the same idea from a completely different direction. TypeSafe AI shipped Jev two days ago, a transformer that outputs calibrated probabilities instead of text, on the argument that human language is the wrong interface for automation. Cua arrived at the same place from the desktop side. When two teams who are not talking to each other both decide the big model should be routed around for the boring 90 percent, that is not a product opinion anymore. That is the shape of next year's cost structure. Repo at github.com/trycua/cua.
Comments