Cua

The detail that decides it is that the drivers work in the background without taking your cursor. Every other approach to desktop automation turns your machine into something you sit and watch, and that single behaviour is the difference between a tool you run during the workday and one you schedule for the night.

Pick it when the target is a native application with no API and no plugin story. Pick Browser Use when everything happens in a browser, because dragging a whole operating system along for a web task is a cost you feel in both time and money.

The unifying API is the selling point and the exposure. Boot times, snapshot semantics, clipboard behaviour, display handling and crash recovery all differ between a virtualised guest and a container, and code written against the shared surface will encounter those differences only in production, on whichever backend the customer chose. Debugging then requires knowing the layer the abstraction was hiding.

Pin one backend per workload and test against that one. What it does right: isolation is the default everywhere rather than a mode you remember to switch on, which for computer use is the only defensible posture.

Building an evaluation harness over three externally authored suites is the correct methodological choice, because it separates the measuring instrument from the thing measured and keeps results comparable to published work by other groups. Screen grounding and full desktop task completion are different competencies, and using suites that isolate each is deliberate rather than accidental.

What the record does not contain is a single number. A harness with no reported results demonstrates good intent and establishes nothing, and the natural suspicion about a vendor-run harness is answered only by publishing runs others can reproduce.

The structure is the familiar one: give away the framework, sell the machines it runs on. That works when the hosted layer is genuinely inconvenient to self-operate, and fleets of virtual desktops qualify, so there is a real business here rather than a hopeful one.

The pressure comes from above. Every lab shipping native computer use compresses the value of an independent harness, and the defensible remainder is the operational work of running desktops at scale. Likely acquirer: an automation incumbent that wants a modern runtime, or a lab that wants the fleet. Position: long the infrastructure, not the abstraction.

Pricing to six decimal places per vCPU-hour tells me this was designed by infrastructure people, and it means our exposure is a function of how long agents sit idle inside running machines rather than how many engineers we have. An agent that stalls does not error, it accrues, and that is the line item I would be explaining.

It runs unattended, which makes it genuinely useful for interface testing in our pipelines. There is no identity federation documented. Approved with conditions: a hard budget cap, timeouts on every sandbox, and a monthly review of the hours.

This is the rare hosted product whose local path is the real one. I ran it under Apple Virtualization on my laptop and under QEMU on the machine in the corner, with no account, no key and nothing reporting anywhere. The driver installs from a shell script I read first.

Exposing an MCP server is the part that made me happy: my own client gets mouse, keyboard, screenshots and a shell on a machine I control, which means I can point any agent I like at a desktop. MIT, so the whole stack is mine to modify. This one I keep.