A Chrome extension for reading hard things. Highlight a passage on any page and ask an AI model to explain it, simplify it, define the jargon, or quiz you on it — without leaving the page.
- Open
chrome://extensionsand turn on Developer mode (top right). - Click Load unpacked and pick this folder.
- The settings page opens on first install — pick a provider, paste its API key, and hit Test.
The key is stored in chrome.storage.local for this browser profile. Requests are made from the
extension's background worker straight to the provider — the key is never exposed to the pages you
browse, and nothing is sent anywhere else.
Every provider here speaks the Anthropic Messages API, so switching one is a base-URL swap. Pick it in settings, or flip between them from the toolbar popup mid-session.
| Provider | Model | $/M in → out | ≈ per question | Questions per $1 |
|---|---|---|---|---|
| Anthropic | Opus 5 | 5 → 25 | $0.0075 | ~130 |
| Anthropic | Sonnet 5 | 2 → 10 | $0.0030 | ~330 |
| Anthropic | Haiku 4.5 | 1 → 5 | $0.0015 | ~670 |
| Z.ai / OpenRouter | GLM-5.3 | 1.40 → 4.40 | $0.0016 | ~630 |
| Z.ai / OpenRouter | GLM-5.3-Flash | 0.075 → 0.25 | $0.00009 | ~11,000 |
| OpenRouter | DeepSeek V4 Flash | 0.065 → 0.18 | $0.00007 | ~14,000 |
| OpenRouter | anything :free |
0 → 0 | $0 | — |
Sized on a typical question (≈500 tokens in, 200 out); the panel shows the real number under every answer. GLM-5.3-Flash is at promotional pricing (50% off) as of writing.
- Anthropic — Claude models, billed per token.
- Z.ai — GLM over
https://api.z.ai/api/anthropic, which speaks the Messages API natively. Key: z.ai. - OpenRouter — one key, ~400 models including free ones. Settings pulls the live catalog with current per-token prices, so the model box autocompletes and the cost readout stays accurate; hit Refresh model list to re-pull. Key: openrouter.ai/keys.
- Custom endpoint — anything else: a gateway, a proxy, a local server. Pick the API shape (Anthropic Messages or OpenAI Chat Completions), enter the base URL, click Grant access to this host so Chrome allows the origin, and type the model name.
Two request shapes are implemented. OpenRouter uses the OpenAI Chat Completions wire because its Anthropic-compatible endpoint is documented as Anthropic-models-only, and the point of OpenRouter here is the cheap non-Anthropic models. Everything else uses the Anthropic Messages wire.
Effort, refusal fallbacks, workspace IDs and reasoning display are Anthropic-specific and are simply
not sent to other providers. On the Anthropic wire with a non-Anthropic provider, the key goes out as
both Authorization: Bearer and x-api-key, since gateways differ on which they read.
There is no way to bill these calls to a Claude.ai Pro/Max subscription — the subscription and the API are separate products. The only subscription route would be a local bridge that shells out to the
claudeCLI; that isn't built here.
Your key isn't scoped to a single workspace, so every request has to say which workspace it acts in. Either fix works:
- Paste your workspace ID into the Workspace ID field in settings. It's the
wrkspc_…value in the ID column of Console → Settings → Workspaces (pasting the whole console URL works — the ID is picked out of it). The extension then sends it as theanthropic-workspace-idheader on every request. - Or create a new key scoped to one specific workspace, and leave the field empty.
A successful Test reports which workspace answered, and fills the field in for you if it was empty.
Select text → a bubble appears with Explain · Simpler · Quiz · Ask….
- Preset actions — the bubble's first three, plus the full set in the panel: Explain, Simpler, Key points, Example, Define terms, Quiz me.
- Your own question — hit Ask… on the bubble (or just type in the panel's input box) and ask whatever you actually want to know: "why does this need ATP?", "how is this different from what the last section said?". The passage and its surrounding page context go along with it.
- Follow up in the same box — the thread keeps the passage in context, and preset actions and typed questions mix freely in one conversation.
- PDFs work too — papers open in the extension's own viewer, where text is selectable. See below.
- Highlights — any passage you ask about is highlighted, and comes back the next time you open the page. Click a highlight to reopen it with the whole conversation you had about it, and keep asking; Alt-click to remove it (and its conversation).
- Notes — every answer is filed in your study notes (
☰in the panel header) as it arrives, so nothing is lost when you close the page. Click Saved ✓ under an answer to take it back out. - Whole page — from the toolbar popup: Summarize this page or Quiz me on the page.
- Right-click any selection for the same actions.
| Shortcut | Action |
|---|---|
⌘⇧E / Ctrl+Shift+E |
Explain the current selection |
⌘⇧U / Ctrl+Shift+U |
Show / hide the panel |
Esc |
Close the panel |
Enter |
Send your question (Shift+Enter for a newline) |
Rebind these at chrome://extensions/shortcuts.
Chrome renders PDFs with a plugin that exposes no text to extensions — you can't select a sentence in an arXiv paper and have anything reach it. So PDFs open in the extension's own viewer instead: pdf.js draws each page to a canvas and lays pdf.js's text layer over it, which is ordinary DOM. Selection, the bubble, the panel and persistent highlights all work exactly as they do on a web page.
Open arxiv.org/pdf/2505.05470 and you land in the viewer automatically — nothing to click. It
handles .pdf URLs and the extensionless /pdf/<id> shape that arXiv and bioRxiv use; for the
latter it checks the response's content type first, so a page that merely lives under /pdf/ is
left alone. Original in the toolbar hands the file back to Chrome's viewer, and the whole
behaviour is one checkbox in settings.
Two things worth knowing:
- Whole-document actions are better here. Summarize page and Quiz me on the page use text extracted from every page, including ones you haven't scrolled to yet.
- Highlights key off the PDF's URL, not the viewer's, so they come back the next time you open the paper — and notes link to the paper, not to a viewer URL.
Local PDFs (file://) also work, but Chrome requires "Allow access to file URLs" for the extension
on chrome://extensions. Scanned PDFs with no text layer can't be selected by anyone — there is no
text in the file to select.
- Provider — Anthropic, Z.ai (GLM), OpenRouter, or a custom endpoint. Keys and model choice are remembered per provider, so switching back and forth costs nothing — cheap model for skimming, Opus for the paragraph that actually matters.
- Workspace ID — Anthropic only, and only for a key that spans several workspaces (see above).
- Model — the panel footer shows tokens and the real cost of each answer.
- Effort — Anthropic only: how hard the model thinks before answering.
lowis snappy,high+ is for genuinely hard passages. Ignored on Haiku. - Explain at this level — keep it simple / student / expert. This changes the tone and depth of every answer.
- Theme — Automatic (follows your system), Light or Dark, for the panel and every extension page.
- Save every answer to my study notes — on by default; turn it off to keep only the answers you file yourself with Save note.
- Open PDFs in the study viewer — on by default; turn it off to leave PDFs to Chrome.
- Surrounding context — how much nearby page text rides along with the passage so the model can resolve pronouns and references. Set to 0 to send the passage alone.
- Show the model's reasoning — Anthropic only: streams a summary of the thinking above each answer.
- Retry declined requests on a fallback model — Anthropic only: server-side refusal fallback for Opus 5. If your account isn't enrolled in that beta the extension drops it automatically on the first 400 and retries; you can also just turn it off.
☰ in the panel (or Notes in the popup) opens the study-notes page: everything you saved,
everything you highlighted, grouped by page, searchable, and exportable as one Markdown file.
A librarian agent can file your saved answers into an Obsidian vault. For each
note it looks at what's already in the vault, puts the answer in the right topic note (creating it if
needed), links related notes with [[wikilinks]], and keeps a source note per page and a
_index.md map of contents.
It only links notes that exist. A concept that deserves a note but has none yet is suggested instead:
it shows as Study next under the note on the study notes page, and in a ## Study next list in
_index.md. When it finishes, any link it wrote to a missing note (which in Obsidian would create an
empty page when clicked) is turned back into plain text. Links you wrote yourself are left alone.
Setup:
- In Obsidian, install the community plugin Local REST API with MCP by Adam Coddington
(GitHub). Several plugins have
similar names;
obsidian://show-plugin?id=obsidian-local-rest-apiopens the right one in Obsidian directly, and the settings page links to it. In its settings, turn on Enable Non-encrypted (HTTP) Server. Its HTTPS port uses a self-signed certificate that Chrome won't trust from an extension. - In Study Buddy's settings, under Obsidian vault, paste the plugin's API key and hit Connect.
- Turn on File every saved answer in the vault, or file notes one at a time from the study notes page.
Under each answer in the panel, an Obsidian line shows what the librarian is doing as it works (Reading _index.md…, Adding to Isaac Newton.md…), then which notes it filed into, with Undo, or why it failed, with Retry. The study notes page shows the same for every note.
Obsidian has to be open while filing. The librarian uses the current provider and key; a separate, cheaper model can be set for it, but very small models are unreliable at tool use.
Page text reaches the model, so the limits are enforced in code rather than left to the prompt: the
librarian can create and rewrite notes only inside its own folder (Study Buddy/ by default), can
only append to notes elsewhere and only if you allow it, can never delete, and stops after 8 writes per
note. Every write records the file's previous text, so Undo restores it exactly, skipping any file
you've edited since. It reads and searches the whole vault to find related notes, so parts of your
notes are sent to the model provider while filing.
manifest.json
background/service-worker.js API calls, both wire formats, SSE streaming, menus, shortcuts
content/content.js selection bubble, panel, highlight engine
content/content.css highlight marks (the only styles in the page's DOM)
viewer/ pdf.js-based PDF viewer (canvas + selectable text layer)
lib/pdfjs/ vendored pdf.js 6.3.289 (Apache-2.0)
lib/config.js providers, settings, model + price tables, prompts
lib/librarian.js Obsidian filing agent: tools, guardrails, undo
lib/vault.js client for the Obsidian Local REST API plugin
lib/markdown.js DOM-building Markdown renderer (no innerHTML)
lib/pages.css shared styles for the extension's own pages
options/ popup/ notes/ settings, toolbar popup, study notes
test/smoke.js end-to-end test
test/librarian.test.mjs librarian against a fake vault and a scripted model
The panel lives in a shadow root, so page CSS can't reach it and its styles can't leak out. Model
output is rendered by building DOM nodes with textContent — never innerHTML — so a page can't
be scripted through an answer.
node test/smoke.js
node test/librarian.test.mjsLaunches headless Chrome with the extension loaded, stubs the Anthropic endpoint inside the
service worker, and drives a real selection → streamed answer → highlight → follow-up → reload
flow, plus the options, notes, and popup pages. No API key or network access needed. Set
CHROME_PATH if Chrome isn't at the macOS default location.
After editing files, hit the reload icon on chrome://extensions and refresh open tabs.
Run the Release workflow from the Actions tab and pick patch, minor, or major. It bumps
the version in manifest.json, uploads the zip to the Chrome Web Store, submits it for review,
then commits the bump, tags vX.Y.Z, and creates a GitHub release with the zip attached. Tick
dry run to only build the zip (downloadable from the run's artifacts).
It needs these repository secrets: CWS_EXTENSION_ID, CWS_CLIENT_ID, CWS_CLIENT_SECRET,
and CWS_REFRESH_TOKEN (npx chrome-webstore-upload-keys walks through creating them).
- Chrome blocks content scripts on
chrome://pages, the Web Store, and other extensions' pages. - Scanned/image-only PDFs have no text layer, so there is nothing to select (OCR would be needed).
- Highlights are matched by their text, so they may not restore on pages whose content changes between visits.
- The extension asks for access to all sites: it already ran a content script everywhere, and the PDF viewer needs to fetch PDF bytes from whatever host serves them.