i'm a CS student and i wanted to organize my course stuff for the start of the semester. I went on Moodle, downloaded all organizational PDFs of each course, all lectures, copied the course descriptions etc and i dumped it all into one folder on my laptop and told Claude Code to organize each one and write comprehensive .md files for each course (what is it about, when is the exam, what should be done across the semester, deadlines, etc) and then after it was done, i told it to build me a system around it.
what i ended up with:
- a fully functional study dashboard on Vercel (React + FastAPI + Supabase) tracking my four courses - lectures, topics, deliverables, exams. fully responsive on mobile, looks decent.
- an MCP server with 40 tools. anything i can do in the UI the AI can do too - create a topic, mark something studied, upload a file, render a PDF as images, whatever.
- plugged that MCP into Claude.ai as a custom connector. full OAuth 2.1 and everything.
- which means those 40 tools are now available in Claude Code on my laptop, the Claude website, AND the Claude iOS app on my phone. i can just open Claude anywhere and it has access to all my course stuff.
- two-way sync between my laptop's `Semester 4/` folder and the bucket. push/pull/watch commands. bucket is the source of truth.
I also made a custom Claude project (in claude.ai) with a system prompt i wrote, so every chat i start in that project already knows my workflow - which courses, how i transcribe lectures, when to mark something as studied vs covered etc.
the part that actually made me post: the MCP that i made has a tool that takes a PDF path and returns the pages as images. I searched some Claude docs and they don't tell you whether it feeds those retrieved images to the model as vision input or just displays them in the UI as attachments when it pulls them through an MCP server. so i tested by asking Claude to describe stuff you can only get by actually looking at the slide and it works.
stuff broke along the way — Vercel's Python runtime doesn't fire ASGI lifespan (FastMCP's task group never starts), FastMCP has a DNS rebinding protection default that's localhost-only, Supabase Storage doesn't like German umlauts in filenames, iOS Safari won't scroll PDFs inside iframes. Claude hit every one of those, figured it out, fixed it.
i never had to say "no do it differently" a SINGLE TIME. this is the first time i've experienced this with Claude Code tbh.
(demo video in the github repo)
i'm a CS student and i wanted to organize my course stuff for the start of the semester. I went on Moodle, downloaded all organizational PDFs of each course, all lectures, copied the course descriptions etc and i dumped it all into one folder on my laptop and told Claude Code to organize each one and write comprehensive .md files for each course (what is it about, when is the exam, what should be done across the semester, deadlines, etc) and then after it was done, i told it to build me a system around it.
what i ended up with:
- a fully functional study dashboard on Vercel (React + FastAPI + Supabase) tracking my four courses - lectures, topics, deliverables, exams. fully responsive on mobile, looks decent. - an MCP server with 40 tools. anything i can do in the UI the AI can do too - create a topic, mark something studied, upload a file, render a PDF as images, whatever. - plugged that MCP into Claude.ai as a custom connector. full OAuth 2.1 and everything. - which means those 40 tools are now available in Claude Code on my laptop, the Claude website, AND the Claude iOS app on my phone. i can just open Claude anywhere and it has access to all my course stuff. - two-way sync between my laptop's `Semester 4/` folder and the bucket. push/pull/watch commands. bucket is the source of truth.
I also made a custom Claude project (in claude.ai) with a system prompt i wrote, so every chat i start in that project already knows my workflow - which courses, how i transcribe lectures, when to mark something as studied vs covered etc.
the part that actually made me post: the MCP that i made has a tool that takes a PDF path and returns the pages as images. I searched some Claude docs and they don't tell you whether it feeds those retrieved images to the model as vision input or just displays them in the UI as attachments when it pulls them through an MCP server. so i tested by asking Claude to describe stuff you can only get by actually looking at the slide and it works.
stuff broke along the way — Vercel's Python runtime doesn't fire ASGI lifespan (FastMCP's task group never starts), FastMCP has a DNS rebinding protection default that's localhost-only, Supabase Storage doesn't like German umlauts in filenames, iOS Safari won't scroll PDFs inside iframes. Claude hit every one of those, figured it out, fixed it.
i never had to say "no do it differently" a SINGLE TIME. this is the first time i've experienced this with Claude Code tbh.
[dead]