Open Source Radar — October 29: DeepTutor Hits 49,000 Stars
Today's board is applications, not skill packs: four self-hostable projects where the agent sits inside the product instead of being bolted onto someone else's. Two of them — a tutor and a cockpit for coding agents — are the fastest movers of the day.
DeepTutor (Python, 48,899 stars) is HKUDS's self-hosted tutoring app, and it added 2,371 stars in the last day, the biggest jump of any AI repo on today's board. Point it at your own PDFs, a course folder or a lecture and it builds a learning path, generates graded quizzes, and keeps a memory of what you have actually mastered rather than what you have merely read. It comes from Hong Kong University's data-intelligence lab, the group behind LightRAG, under Apache 2.0 — so it runs on your own API key or a local model. That last part is the use case: a study plan over material you already own, or a tutoring layer in a school that will not ship student data to a vendor.
OpenBot (TypeScript, 3,524 stars) is a desktop app that runs several coding agents side by side, each with its own workspace and conversation: Codex and Claude Code natively, Grok, Gemini, Cursor, Cline and OpenCode through the agent-client protocol. It gained 1,054 stars today. Queues survive a crash, agents can message each other and hand off files, and there is an embedded browser they can drive. Two caveats come straight from its own README: it is a development preview that runs agents with full disk access and no per-action approval, and it is licensed PolyForm Noncommercial — source-available, not open source in the OSI sense. Use it as a cockpit for unattended work; do not put it on a machine you lack backups for.
Macro (Rust, 5,614 stars) is a whole team workspace — email, chat, docs, tasks, calls, CRM — that a New York and Toronto team built in-house and ran its own fifteen-person company on for two years before it trended today. Everything cross-links in one graph: a support email becomes a task, the task links the pull request, and agents read the same graph the humans do. It is AGPL-3.0 rather than open core, with a Docker stack for self-hosting. The bet, stated plainly in the README, is that agents are only useful when mail, chat and CRM live in one system instead of being stapled together with MCP and Zapier. The cost is migration: 7,557 commits of opinions about how work software should look, and 344 open issues.
RuView (Rust, 101,423 stars) turns cheap WiFi into a sensor. A roughly $9 ESP32 board reads the radio reflections off the people in a room and reports presence, breathing rate, heart rate and falls with no camera and no wearable, according to the project, and it exposes itself to Home Assistant, Apple Home and Alexa through Matter. What makes it worth a second look rather than a skeptical scroll is that the repo publishes its misses: the presence encoder sits at 82.3% on a held-out test after the older "100% presence" figure was retracted, and the bundled on-device pose model scores 3.0% against its own 35% target, with only the separate MM-Fi benchmark model reaching 82.69%. Use it for occupancy or elder-care alerts where a camera is not allowed. Do not buy it for pose estimation yet.
Worth watching this week: whether DeepTutor holds its lead once the trending list turns over. Forty-nine thousand stars for a tutor in a market where the big chatbots answer homework for free is a signal about self-hosting, not about tutoring.
Would you hand a desktop agent full disk access to get work done unattended? Write to us via our contact page.
