Raphael Amorim has not decided that open source was a mistake.

He is still maintaining Rio, his public MIT-licensed terminal emulator.2

Canario, the terminal he built around agentic coding work, is going closed source.1

His reason is less ideological than practical: he no longer wants to pay the human maintenance cost of keeping this particular project open.

Free code creates paid work in attention

In his August 11 post, Amorim says he has been doing open source since he was 17 and has repeatedly helped companies publish internal projects. Rio remains open.1

Canario is different. He built it for himself, without VC money and without a plan to turn it into a large business. Rio and other work already consume his available time and energy.1

Publishing source code is only the beginning of an open repository.

Someone has to read issues, inspect pull requests, answer discussions, distinguish bugs from preferences, reject changes and then explain why they were rejected.

That is the cost Amorim describes.1

AI makes the equation unusually asymmetric.

A pull request is cheaper to generate than to review

Amorim explicitly points to low-effort issues, generated pull requests and AI-enabled spam as part of the maintenance burden.1

He does not publish a count. It would be wrong to claim Canario was flooded by thousands of AI contributions, or that AI was the only reason for closing the source.

The narrower point is enough: generating a low-effort contribution is getting extremely cheap while reviewing it remains human work.1

For the contributor, an agent can produce a patch in minutes. For a maintainer, the cost is not how many keystrokes produced the diff. The maintainer has to understand the intent, review the change, check tests and side effects, and decide whether the project should own that behavior for years.

A contribution can be nearly free to produce and still expensive to absorb.

Closing the door instead of building a better bouncer

There are other possible responses: stricter templates, more CI, automated triage, contribution scoring, perhaps even an agent reviewing the agent.

Amorim chooses the simpler boundary. Canario will move at his own pace, without drive-by requests, spam issues or generated PRs that he feels obliged to process.1

That decision also removes some benefits of open development: outside fixes, current-source inspection and unexpected contributions. He does not claim otherwise. For this project, he has decided the exchange is not worth the attention it consumes.

According to Amorim, older Canario source remains in older Rio versions and can still be forked.1 Current Rio remains public, MIT-licensed and actively maintained, with more than 7,000 GitHub stars when we checked.2

This is not a blanket conversion to closed software. It is a maintainer deciding which room he still wants to clean after everybody leaves.

Open source is also a social queue

Open-source discussions often center on the license and availability of code.

For maintainers, a public repository is also an inbox.

Bugs, feature requests, opinions, patches and support questions arrive there. Now some contributions can arrive before their authors have deeply understood, or even fully read, what was generated.

A license decides what other people are allowed to do with code. It does not decide how much attention the maintainer owes them.

That distinction becomes more visible as code generation scales. Communities must increase review capacity, make contribution more expensive, accept more noise, or close some entry points.

Canario illustrates the last option.

Abundant code makes evidence more valuable

Agents can genuinely make open-source contribution more accessible. A person who understands a project but writes slowly, is learning a language, or needs help with repetitive work can produce a better patch with an agent.

The same lower barrier also makes it easier to submit code before understanding the project well enough.

Both can be true.

So “AI code is bad” is not a useful rule. The scarce signal becomes evidence that a contribution deserves another person's time: a clear reproduction, tests, a small scope, motivation, and visible understanding of the surrounding code.

Those signals mattered before agents. They matter more when 300 lines no longer imply 300 lines of human effort.

Amorim has chosen not to solve that filtering problem inside Canario.1

A technology that makes construction cheaper can make selection more important. Sometimes the resource worth protecting is the one it did not automate: maintainer attention.