To discuss a text with an AI in a chat, you copy the passage, wait for the answer, then copy the next passage. The conversation grows into a long thread, while each remark belongs to a specific spot in the text, and you end up losing track of which paragraph an answer was about.

Commentaire tries the other layout. You read a document, hold down a key to speak, and the remark attaches itself in the margin, level with the passage you were reading. An AI co-reader stays silent while you read and answers in the same margin when you ask it a question.

The Commentaire app: a text on the left, and in the margin a comment and a question attached to their paragraphs, with Opus's answer
The app's interface, in French, loaded from its public code with the repository's test essay. With no compiled engine, the remarks and Opus's answer are sent by hand, in the engine's format.

How a remark travels

Key held down→Whisper transcribes on the Mac→Haiku sorts→Opus answers questions→Margin

Whisper transcribes the voice on the Mac. Claude Haiku sorts each segment into reading, comment, question or instruction, and Claude Opus writes the answers to questions, which stream into place. Claude runs from the command line, on the user's subscription, with no API key and no server. That choice dates from the day after the spec: written on the evening of April 15 for a web app in Next.js, it was replaced on the 16th at 1:40 p.m. by a desktop app architecture in Tauri, for lack of an API key and a server, on a Mac with 8 GB of memory.

The full chain was running on April 16 at 8:40 p.m., and testing took place at dawn the next day.

MeasurementsApril 17, 2026
Remarks sorted correctlyby Haiku 6 / 6
Memory usedceiling set at 1.2 GB 454 MB
The black mark shows the ceiling that was set. The same measurements pointed to the next piece of work: the response time of Claude, called as a subprocess.

Two fixes along the way

At first, recording started whenever the volume went over a threshold. Whisper ended up transcribing *Bruit de porte* (a door noise). On the morning of April 17, a push-to-talk key replaced the detection.

Claude Code had passed the classifier's instructions through an environment variable that the claude command ignores. For about ten minutes, the classifier replied as if it were in a conversation. The note left after the fix: always read the command's help before guessing.

Two agents that don't talk to each other

The project was also a way to get Cowork and Claude Code working together, with no conversation between them, through the repository and the tickets.

Coworkwrites the spec and the architecture into the repository, and the tickets into Linear.
Linearone ticket, one branch with the same name. Claude Code takes the first free ticket and moves it to in progress.
Claude Codepushes at every step, rereads an instructions file at every commit, and closes with "done, not done, problems, test command."

On April 17, Claude Code noted that three new ideas (keeping a history, versioning, letting the AI edit the text) fell outside the planned scope, and that the architecture would have to be rethought to take them on.

The app works end to end, from voice to the answer in the margin, and its code is public.