Summarize without opening the popup. Right-click any page and pick Summarize this page, or use the new keyboard shortcut (default Alt+Shift+U, remappable at chrome://extensions/shortcuts). A system notification lets you know when it's ready, and if you open the popup while it's still generating, you'll see the normal loading view and live result instead of a blank Home screen.

Check the summary against the source. Click any bullet (or line) in a summary and Apogee scrolls to and highlights the passage of the original page it's most likely grounded in, using the same on-device retrieval that powers Ask. A quick way to spot-check that nothing's been invented. (Chromium browsers only for now.)

Copy as Markdown, a new button next to the existing copy button, exports a formatted note (title, source link, and summary body) ready to paste into your notes app of choice.

Cancel button for Summarize and Ask that actually interrupts generation server-side, instead of just hiding the UI while the job kept running.

Past Summaries on Home, with page titles, so reopening the popup shows your recent summaries and you can jump back into any of them with one click.

Everything else

• Rotating loading text while summarizing, a refreshed icon set, better contrast in light theme, and a round of smaller UI/spacing polish.
• Reliability: summaries now get saved by the background job itself the moment it finishes, not by whichever popup happens to still be open — this also fixes a bug where a summary could be silently lost if you closed the popup too early.
• Security: tightened the Ollama host allowlist, narrowed the YouTube caption-URL allowlist, and added Telegram/Slack/Discord/Microsoft Teams to the list of sites whose content is never saved to disk.
• Several bug fixes: PDF text no longer gets lost when asking a follow-up question, lingering progress banners, copy buttons not reappearing after reopening the popup, and more.
• Removed dead code and an unused font; npm audit is now clean.

Full details: CHANGELOG.md (https://github.com/darshi1337/apogee/blob/main/CHANGELOG.md#018---2026-07-21)

Full Changelog: https://github.com/darshi1337/apogee/compare/v0.1.7...v0.1.8

Added
- In-browser AI on Firefox via Transformers.js — Firefox lacks the browser.offscreen API so WebLLM/WebGPU can't run there; Firefox now gets its own on-device provider running ONNX models via WebAssembly. Ships with SmolLM2 360M (default), Qwen 2.5 0.5B, and Llama 3.2 1B. On slower hardware, Local Ollama remains the faster option.
- Retrieval-augmented "Ask" answers — long pages are now embedded locally and answered from the most relevant passages instead of truncating to the first ~8000 characters, so questions about content deep in long articles, PDFs, or transcripts work correctly. Falls back to truncation if embedding is unavailable (always on Firefox).
- Live Ollama model list — settings now show the models you've actually pulled (via /api/tags), falling back to the hardcoded list when Ollama isn't reachable.
- Per-model chunk sizing for summarization — chunk size for Ollama models scales with the model's context window, so capable models need fewer passes.

Fixed
- SponsorBlock lookups now work on the Firefox build (sponsor.ajay.app was missing from the Firefox connect-src).
- The Transformers.js engine no longer leaks WASM memory (hundreds of MB) when reloading after a failure.
- A failed suggested-questions job no longer permanently blocks prompt regeneration for that page.
- Cached view state no longer stores raw page URLs on disk — only a hash (URLs can carry session tokens).
- A stale/unknown provider setting (e.g. carried over from the other browser's build) falls back to this build's in-browser provider.
- README privacy section now discloses the SponsorBlock lookup and the Firefox WASM path; "Connected" status reflects actual WASM availability.

Security
- The Ollama status/model-list probe now enforces the same loopback-only host validation as every other Ollama request, closing the last gap in the extension's SSRF rule.

Local Ollama mode no longer needs a separate backend server. The extension now talks directly to Ollama's HTTP API (http://127.0.0.1:11434) from its background service worker — no more installing/running the apogee-backend Node process. See the README's Advanced: Local Ollama Mode (https://github.com/darshi1337/apogee#advanced-local-ollama-mode) section for the one-time OLLAMA_ORIGINS setup this requires.

Changed

• Local Ollama mode connects directly to Ollama, cutting out the local Node relay entirely.
• PDF summarization now extracts text fully client-side via pdf.js, and works in both WebLLM and Local Ollama modes (previously Local Ollama only, via the backend).

Removed

• The apogee-backend Node.js package — fully superseded by the direct connection above.

Added

• ESLint + Prettier tooling (npm run lint / npm run format).
• GitHub Actions CI: format check, lint, tests, and build run on every push/PR.
• CONTRIBUTING.md and CHANGELOG.md, issue/PR templates.

Internal

• Repo cleanup: removed stray build artifacts from root, simplified the root package.json, tidied .gitignore, moved README-only logo assets into .github/assets/.

Full Changelog: https://github.com/darshi1337/apogee/blob/main/CHANGELOG.md#016---2026-07-17

The apogee-backend server has been completely rewritten from Python/FastAPI to Node/Express. If you're using Local Ollama mode, you'll need to reinstall the backend:

cd apogee-backend
npm install
npm install -g . # optional, exposes the `apogee` command
apogee setup
apogee doctor
apogee


Python, pip, pyproject.toml, and requirements.txt are no longer part of the project. Apogee now only requires Node.js 20+. The HTTP API, CLI commands (setup/doctor), and the default 127.0.0.1:8000 endpoint remain unchanged, so the extension itself requires no reconfiguration beyond reinstalling the backend.

• Fixed an issue where a dropped connection during summarization (for example, when Chrome suspended the extension's background service worker after ~30 seconds of inactivity) could be incorrectly cached as a completed summary, resulting in a truncated response with no visible error.
• Fixed a related issue where reopening the popup after a failed stream would repeatedly retry the same dead connection instead of recovering normally.
• Replaced idle timers with chrome.alarms, ensuring the in-browser model is unloaded reliably even when the background service worker is suspended, preventing unnecessary GPU and RAM usage.
• Fixed a race condition that could trigger "Only a single offscreen document may be created" when multiple actions initialized the model simultaneously.
• Fixed an issue where the in-browser model sometimes failed to unload after a background-worker restart.
• Suggested-question parsing now preserves legitimate leading digits (for example, "2025 predictions?").
• Fixed two issues in the YouTube description cleaner: one overly broad regex removed legitimate content, and hashtag filtering now removes only hashtag-only lines instead of entire sentences mentioning hashtags.
• The popup version number is now read directly from the extension manifest instead of relying on a hardcoded value.

• OLLAMA_HOST is now fully respected, allowing Ollama to run on custom hosts and ports.
• Remote PDF downloads are now limited to 50 MB with a 30-second timeout, preventing unbounded downloads.
• Fixed an issue where long summaries could trigger significantly more model calls than intended on CPU-only systems.
• Fixed file:// PDF URLs containing page fragments (such as #page=2) being incorrectly rejected.
• Numbered bullet points are now preserved correctly in multi-part summaries.
• Improved environment-variable validation, including clean handling of invalid APOGEE_PORT values instead of crashing.
• Removed a startup hook that could accidentally terminate unrelated processes listening on port 8000.

Privacy, YouTube, and reliability improvements.

• Privacy — "Save history" toggle, "Clear cached data" button, sensitive sites (Gmail, Outlook, Proton, etc.) never cached, and URL‑hashed storage keys.
• YouTube — transcripts now actually work (no more summaries hallucinated from the title); marketing boilerplate stripped from descriptions.
• Reliability — summaries, answers, and prompts keep running in the background if the popup closes, and the popup reopens to your last view.
• UI — cleaner progress box and Clear‑data button.
• Security — SSRF/CORS hardening and a stricter extension↔backend boundary.

Install: load the extension zip below (chrome://extensions → Load unpacked), or pip install apogee_browser-0.1.4-py3-none-any.whl for the optional local Ollama backend.

Full changelog: https://github.com/darshi1337/apogee/compare/v0.1.3...v0.1.4

• perf(sw): Keep offscreen document alive with a 5-minute idle timeout. Previously, the offscreen document (including the loaded MLCEngine and WebGPU device) was torn down on every focus loss of the action popup, triggering full model reloads on subsequent interactions.
• fix(webllm): Add content chunking and truncation (lib/chunk.js) and implement map-reduce summarization to match backend behavior, preventing context overflows on long articles in small models.
• perf(popup): Reuse extracted page data for follow-up questions instead of re-running the full Readability DOM clone.
• perf(popup): Return the DOM extraction result directly from executeScript, removing 3 extra script injections that mutated the host page. Unused run_extraction.js has been deleted.
• fix(popup): Correctly handle WebGPU check failures and avoid caching fail-open WebGPU results so "WebGPU not supported" warnings are correctly displayed if a transient messaging error occurs.
• fix(popup): Bound the summary cache with FIFO eviction (capped at 50 pages) to prevent unbound memory growth.
• fix: General hardening including UUID stream IDs and escaping quotes in escapeHtml.

• fix(backend): Accept -, *, and • interchangeably as bullet markers in summaries to prevent dropping chunks whose bullets differed from •.

Hardened privacy: Scoped permissions strictly to the active tab and locked network egress to 127.0.0.1.
Improved streaming UX: Added support for smooth token-by-token response rendering in the popup.
Version alignment: Bumped extension version to 0.1.2 and updated internal dependencies to keep Firefox and Chrome parity.