StudyLoop

Inspiration

Here is the thought that hooked me. Normally, turning an idea like this into a real product is a mountain of work: build a mobile app or a web app, then wire an AI into it — API keys or your own server, optimization, running costs — all before a single student benefits. StudyLoop needed none of that. The student opens the ChatGPT desktop app, opens the built-in browser, and the whole thing is run by their own AI — the one they already talk to every day, maybe for months, maybe for years. Nothing new to install. No habit to break. I was a teacher and a school principal for 16 years, and right now I'm a student again, preparing for my own university exams — so I know both sides. Students don't need another app; they need the tools they already use to work together.

What it does

Open StudyLoop in ChatGPT's browser and ask for a quiz on anything. In about ten seconds, your agent builds a session live on the page through WebMCP tools: question cards deal in, a study plan appears. You answer on the page — the grading is done by the page itself, deterministically, not by the AI's mood. Then your agent reads your results through a state tool — misses, per-topic accuracy, cards you flagged with "I don't get this" — and writes back: explanations attach to the cards you missed (they glow), and new questions arrive targeting your weakest topics. Mastery bars, streaks, and a session summary close the loop.

Because your own agent generates all the content, StudyLoop works in any language — I tested it in Turkish, in my own casual speaking style, and the loop held. And it degrades gracefully: with no agent present, it's still a fully working manual quiz builder. The header badge honestly reports the state: agent-ready, tools blocked, or manual mode.

How we built it

Vanilla JavaScript on the W3C WebMCP spec — seven tools registered via document.modelContext.registerTool, each description written as a prompt for the agent (usage example, failure modes, safe defaults), with domain errors returned as {error: "...what to do instead"} so agents self-correct. There is no AI backend, no API keys, no server-side intelligence — the student's agent IS the brain; the site provides structure and shared state. A dev-only simulator (?dev=1) drives the page through the standard getTools()/executeTool() path for testing in any browser. Hosted on Firebase. Built in one day of intensive pair-work with Claude Code, including two adversarial multi-agent code reviews.

Challenges we ran into

My biggest fear was simple: would the model even call the tools? The starting point is just a prompt — what if nothing happens? And could I trust it to handle right and wrong properly? That fear shaped the design: tool descriptions are written like instructions to the agent, malformed question batches are rejected before they can touch a session, and grading belongs to the page, not the model — the agent writes the questions, the page scores the clicks. Then the first real test came: I typed one sentence into ChatGPT, and ten seconds later the cards were on the page. That moment answered everything.

The code reviews caught things I'm glad judges will never see. The best one: my state tool claimed the spec's readOnlyHint while every call quietly wrote an activity entry — a chatty, polling agent could have pushed the student's real history out of the feed. Reads now leave one honest, coalescing trace. Live QA taught me another lesson: a bare "check how I'm doing" produced a chat-only summary with no page write-back. Agents do what you ask — so the prompts below ask explicitly. And honestly? Using GitHub for the first time scared me a little too. The repo is public now, MIT licensed.

What we learned

The bigger lesson: I didn't need to build a mobile app at all. Open the browser and it's done. If every website worked in this harmony with agents, the web would be a completely different experience.

What's next for StudyLoop

I'll use it myself for my own exams first. Then: spaced repetition across sessions, shareable study sets, and a class mode where a teacher's agent maintains study plans for a whole group.

Built With

• learning
• study