Claude computer use in production: isolate the desktop, log the clicks

Computer use lets Claude operate a GUI: browser, desktop, internal thick clients that have no API. It is powerful and easy to romanticise. In production it is an interface of last resort. If Salesforce has an API, use MCP or tools. Do not click the UI because a demo looked like a human.

When the system of record truly has no API, computer use can still be the right seam. Treat it like RPA with a language model: sandbox, budget, logs, human review.

Production constraints

• Dedicated VM or container, not an engineer's laptop
• Network allow-list
• Step and time ceilings, same as any agent loop
• Screenshots and actions in the trace
• A human on submissions that spend money or change access. See human-in-the-loop.

Evals are screenshots plus outcomes

A golden set for computer use includes the start URL, the goal, and the expected end state. "It usually finds the button" is not a case. Pair with CI gates even if the runner is heavier than a JSON eval.

Prefer tools as soon as they exist

Every month of GUI automation is a month you should be asking for an API. Computer use can unblock a process. It should not become the architecture.

If this is the only way through a legacy screen, we will say so. If it is a shortcut around a perfectly good API, we will say that too. That honesty belongs in week zero of scope.