Hey Product Hunt 👋 We’re launching Cobalt—an AI developer with its own cloud computer. The idea behind Cobalt is simple: coding agents don’t only need better models. They need somewhere real to work. Real software development involves more than editing files. It requires dependencies, databases, services, Docker, tests, credentials, a running application, and enough time to complete the job. Much of that still falls back on the developer to configure and supervise locally. And git worktrees are not the solution, they simply add to the confusion! With Cobalt, you connect a GitHub repository and describe the outcome you want. Cobalt creates an isolated cloud computer, works directly in the repository, runs the tests, starts the application, and shows you the result through a live browser preview. It keeps working after you close the tab and returns the finished change as commits and a pull request. Each task has its own computer, so features, bug fixes, prototypes, and experiments can run in parallel without additional local checkouts or worktrees. We initially thought the main problem was giving a coding agent remote compute. This is what Cursor, Codex and ClaudeCode can already do. But it turned out that what mattered just as much was persistence and the feedback loop: letting the agent run the complete application, letting you inspect the result while it works, and returning something reviewable rather than a collection of suggested edits. If you already use coding agents on real repositories, I’d particularly value hearing where your current workflow still breaks down. Thanks for taking a look at Cobalt. I’ll be around throughout the launch to answer anything.

Cobalt - Give your coding agent its own computer

Replies