DeepTutor's 2,282-Star Day Measures the Pitch, Not the Tutor
An open-source agent-native tutor with workspaces, mastery paths and practice, shipped as one container or three pip lines; the star spike says more about the pitch than the software.
One container, one named volume, port 3782 bound to localhost. That is the entire published install for DeepTutor, and the repo picked up 2,282 stars in the last day on a three-word pitch: lifelong personalized tutoring. The all-time counter sits at about 3, so essentially every one of them landed at once, on the strength of that phrase and nothing else I can see in the repo.
One level down the pitch grows: separate workspaces for subjects and projects, Books, Mastery Paths, Reading and Practice with chat tied to whichever activity is open, personas and skills carried across workspaces, account-level companions carrying reusable instructions of their own. The site calls it an “Agent-native Learning Companion that grows with you,” and promises one agent loop over your tools, skills and sources.
The published image is one command; the source route wants Python 3.11–3.14 and Node 22 LTS, and the headless CLI arrives only after 1.7.0. Editor’s note: that is read off the repo and the project site, not from a running install. What I doubt is the thing the feature list skips — one loop deciding what to do with a study plan that has gone sideways.
The self-hosted default is why I have not written this off, but whether that loop is worth a term is unanswered. Skip it if knowledge bases are the point and you would rather not set up an embedding model. Otherwise: one subject, notes written by hand, an agent answering only from those — run the container on 127.0.0.1:3782 and hand it that before it gets a curriculum.