A study app on your own computer: courses, written lessons, checks of what stuck, notes, whiteboards and a daily plan.
No account. It uses your own DeepSeek or OpenRouter key.
- Chat. Ask about your notes and files, with citations. It can make notes, cards and drawings, and it asks you to try graded work first.
- Courses from a chat. Say what you want to learn. You review a draft path, then create it.
- Lessons. The model says what kind of knowledge a topic is, and the code picks the formats it must be taught in (19 block types). You can edit, rewrite a part, regenerate and restore history.
- Build from a folder. Point it at your slides and past papers. The Checker re-solves every auto-marked answer against your files, throws out what they contradict, and leaves its report in the course. One measured run: one PDF, 5 lessons, 165 seconds, $0.04 on DeepSeek.
- Marking. Exact answers are marked by code. Written answers are marked by your model against a rubric. If it can't be marked, it says so and scores nothing.
- Checks. Questions from memory, days after you studied. Only those count toward readiness, and readiness is compared with your real grade.
- Today, a planner, notes, whiteboards, flashcards. These work offline.
- Reads from you: pasted announcements, a syllabus or an exam timetable PDF become suggested dates you confirm.
- Spending limits: $1 a course and $2 a day by default. You can change them in Settings, which also shows what each thing cost.
These are on bilimind.com, the hosted version: sync across devices, mobile, classes, Moodle, timetable links, narration and speaking practice, search across everything, backups, and AI with no setup.
Your key's text models can't read photos, so a timetable photo is refused with the reason. Use a PDF instead.
Tested on Linux, as a package, end to end. Windows and macOS builds are set up in the release workflow, but this version hasn't been run on them, and the builds aren't code-signed.
Your data is in a folder on your computer, and your key is sealed with the system keychain where there is one. What leaves your computer, and only when you use it:
- The text a model reads goes to your AI provider: your question and the excerpts of your files it needs.
- A link you add is fetched from its site. Private and local addresses are refused.
- The code sandbox downloads its Python and SQL runtimes from cdn.jsdelivr.net the first time an exercise needs them.
There are no analytics and no account.
Node 22 or newer and pnpm 10.
git clone https://github.com/bilimind/desktop.git
cd desktop
corepack enable
pnpm install
pnpm start # add ELECTRON_DISABLE_SANDBOX=1 on Ubuntu 24.04+
pnpm test
pnpm --filter @bilimind/companion e2e # after spa and build; nothing shows on screen
pnpm dist # installers for the OS you are on, in apps/companion/release/apps/companion is the app, packages/* the engine, types and offline player, apps/cli the command line (it saves a course as one offline page, and its narration sends the words to Microsoft's online voice service), and skills/bilimind a skill built on it. More: docs/DEVELOPERS.md.
AGPL-3.0. Sample courses: showcase/courses.