Security
Effective 2026-10-23
Architecture
- A Next.js app on Vercel with a Supabase Postgres database. Your browser talks only to this site. The chat runs on the server, which calls OpenAI's API with
store: false; the OpenAI key never reaches the browser. Sign-in, when available, is optional: Google or an emailed link through Supabase Auth, plus the feature-flagged Sign in with ChatGPT. - If switched on, Sign in with ChatGPT uses OpenAI's OAuth with PKCE. Tokens are stored AES-256-GCM encrypted, keyed by a hash of an httpOnly session cookie, and only the server calls OpenAI with them.
- The catalog is refreshed nightly by a server-side job using a service-role key that never reaches the browser.
- The live SIS and Stellic tools run only on your own Mac, through your own signed-in Chrome. The hosted connector never has them, and no SIS or Stellic session leaves your machine.
Database access
- Row-level security is enabled on every table. A database check fails if any table in the public schema lacks it.
- Account tables (
chats,chat_messages,plan_versions,student_profiles) have owner-only policies: a signed-in student can read and change only their own rows. Course history lives only instudent_profiles, and only after an opt-in.chat_usage(hashed IP, token counts, cost; purged after 30 days) is service role only. shared_plans: no public policies; reachable only through server code that requires the 80-bit plan code. The site sends the code in a request header rather than the URL, so it stays out of request logs. Each plan is capped at 256 KB, and plans with no save in 180 days are deleted by a daily job.- Catalog tables (terms, sections, ratings) are public read and writable only by the service role. The sync log is service role only.
Abuse limits
The chat is limited per client and per day, with a daily spend ceiling. Each client without an account (by hashed IP) and each account may also use at most a share of that ceiling (15% by default), checked atomically in the database, so a few clients can't spend the whole day's budget. The plan, catalog, audit-parsing and connector endpoints are rate limited per client IP and cap request sizes, and creating new plan codes has its own much lower hourly limit. Those per-IP rate limits are best effort and enforced per server instance; the daily chat limits are enforced in the database.
Cart changes need your approval
The live tools can add or remove items in your SIS shopping cart only after a native macOS dialog on your own screen lists the exact class numbers and you click approve. The model cannot press that button, and cart changes are refused on other operating systems. Nothing ever clicks Enroll.
Browser protections
The site sends a Content Security Policy, HSTS, frame denial, no-sniff, a strict referrer policy, and a restrictive permissions policy.
Reporting a vulnerability
Open an issue on GitHub saying you have a security report (leave the details out), and you'll get a private way to send them. See security.txt. Please give us reasonable time to fix before disclosing, avoid accessing other people's data, and do not degrade the service. There is no bug bounty, but we will credit reporters who want it.