BROWSER-ONLY · 0.8 · 19 OCTOBER 2026

A browser notebook. AI by choice.

No installation and no API key are needed to write, learn and keep your work. Optional AI is a separate choice.

1 / Open the notebook once online

Open JotHarbour and wait for “Offline files ready”. Keep that browser profile and site storage. Writing, local reference search, solution sheets and the fixed STEM labs then work without an internet connection. Save a bookmark for returning later.

Offline access depends on your browser successfully caching the app. Private browsing, storage cleanup or browser eviction can remove it. The first visit always needs internet. Keep verified notebook exports as backups.

2 / Optional browser AI

Open AI options → Browser AI. Choose a model, check your browser, agree to the model cache, then prepare it. The model runs in a Web Worker using WebGPU, with no publisher AI server or key. After a successful setup, the same model can work offline while its files remain cached. Load it again after reopening the tab. After app updates, reconnect once to prepare the current runtime before relying on offline AI.

These MLC conversions use pinned model revisions and per-file SHA-256 integrity metadata. WASM runtime code is served from this website. Model files come from Hugging Face after your agreement, which can use mobile data. Models are Apache 2.0. These memory figures are estimates, not guarantees.

WebGPU support varies by browser, graphics drivers, GPU features and device memory. Many phones cannot run these models. Notebook tools and fixed labs continue to work without them. No physical-device WebGPU inference test has been completed for this release.

Both starters are text-only. They do not recognise handwriting or analyse a drawing. Small models can be inaccurate or unsuitable. Use the fixed circuit, projectile, quadratic, wave and sensor labs for idealised calculations and check your course materials. Models have no tools, code execution or live web browsing.

Use your existing AI website without an API key

Copy your question from the notebook and use the ChatGPT or Claude link in AI options. Paste the question there yourself, using your own account or subscription. No notebook content is sent by those links. Paste any answer you want to keep into a solution-sheet note. This is separate from automated API access.

3 / Cloud access stays advanced and locked

Normal use has no API-key system. If you deliberately open Advanced cloud connection, you must acknowledge browser-key risks, choose OpenAI or Claude, and supply a key you own. There is no publisher API key, proxy or included API credit. Provider API charges and policies apply; chat subscriptions are separate.

Use a dedicated restricted key with a low spending limit. Prefer temporary use. Optional remembered keys are encrypted locally with your own passphrase. Hiding the tab locks access; unlocking expires after five minutes. Each question asks before sending recent text, matching reference excerpts and optionally the current page image. Browser CORS or provider policy may block access; the app never bypasses this with a hidden proxy.

4 / Understand the vault’s limits

Saved keys use AES-256-GCM authenticated encryption, PBKDF2-HMAC-SHA-256 with 600,000 iterations, a fresh random 16-byte salt and 12-byte nonce. Your passphrase and derived encryption key are not saved. API keys are excluded from notebook exports. Forgot your passphrase? Delete the encrypted credential and enter a new key.

Hashing an API key alone would prevent using it. The salted derivation protects the passphrase; authenticated encryption protects the recoverable credential. Once unlocked, the key must be available to browser code. Encryption cannot stop malicious extensions, compromised site code or an attacker using an unlocked session. OpenAI recommends avoiding browser API keys. The credential vault does not encrypt notebook text, model caches or exported files.