Let me deal with the confusion first, because I lost twenty minutes to it and you should not have to. Roo Code is the VS Code extension that a lot of people know. Roomote is a different product from the same team, a cloud agent rather than an editor extension, and roocode.com now redirects to roomote.dev, which strongly implies where the company's attention has gone. If you came looking for the extension and landed on a cloud platform with a $499 tier, that is why.
I do not love this. Redirecting the domain of a well-known open source tool to a commercial product blurs a line that ought to stay sharp, and the goodwill the extension earned is doing work for a product with a different licence. Now the product, which is better than my grumbling about the naming suggests. The core idea is that most coding agents fail because they cannot check themselves.
A model writes a patch, cannot run it, cannot see the test output, cannot observe that the dev server now throws on startup, and so hands you confident-looking code that does not work. Roomote runs your actual development environment, which means the agent can execute, observe, and iterate before it shows you anything. This is the right architecture and it is the direction the whole category is moving. The second idea is more interesting and I have not seen it done this deliberately elsewhere.
After the agent writes the code, a different model reviews it. Not the same model asked to check its own work, which mostly produces agreeable self-assessment, but a separate model with separate failure modes. In practice this catches a meaningful class of error, because the mistakes a model makes when generating are not the mistakes it makes when critiquing, and using two of them creates something a little like a review. It is not as good as a human who knows the system.
It is considerably better than nothing and better than self-review. Pull requests come back with live previews attached, so approving is a matter of looking at the running thing rather than reading a diff and imagining. For frontend work particularly, this collapses the review loop. The chat integrations across Slack, Teams, Discord and Telegram are the feature I am most ambivalent about.
Being able to fire a task at an agent from Slack is convenient and it also encourages exactly the wrong behaviour, which is dispatching underspecified work and hoping. The quality of what comes back tracks the quality of what you sent, and a one-line Slack message is rarely enough. On the licence, be precise. Fair Core License 1.0, converting to Apache 2.0 after two years.
You can read it, modify it, self-host it, and you cannot build a competing commercial product on it. That is source-available. It is a defensible commercial choice and it is not open source, and the distinction matters if you are making a decision that depends on it. Pricing runs $49 to $499 monthly for cloud depending on seats, with a seven day trial, and self-hosting is free up to ten users with paid licences from $249 for larger teams.
My recommendation is to self-host and evaluate on your own repository, because agent quality varies enormously by codebase and no demo tells you how it performs on yours. Three point seven. Good architecture, a genuinely smart review step, a licence that needs reading and a naming situation that needs fixing.