Developer builds local-LLM browser extension after Mozilla kills Orbit

Apogee is a browser extension designed to summarize articles, videos, PDFs, and emails using AI models that run entirely on the user's device, with no external servers or API keys required. The project was born from frustration with Mozilla's discontinued Orbit project, which attempted similar functionality but relied on centralized Mistral 7B API servers and cached summaries server-side—creating both privacy concerns and the structural vulnerability that led to Orbit's shutdown.

The extension offers two execution modes. In-browser AI uses quantized language models that run directly in the browser: on Chrome, Edge, and other Chromium-based browsers, it leverages WebGPU (via WebLLM) to execute models on the GPU; on Firefox, it uses Transformers.js to run smaller ONNX models on the CPU via WebAssembly. Available in-browser models include Qwen 2.5 1.5B (~900 MB, default on Chrome for multilingual summarization), SmolLM2 1.7B (~1 GB), Llama 3.2 1B (~700 MB), and Phi 3.5 Mini (~2.2 GB). Firefox defaults to SmolLM2 360M (~270 MB), the smallest and fastest option for CPU execution. On first use, model weights download automatically and cache locally; after that, everything runs offline without any external calls.

For users who want larger, more capable models (4B to 8B+), Apogee includes Local Ollama mode, which connects directly to a local Ollama instance over 127.0.0.1. Supported models include Gemma 3 (~4B), Qwen 3 8B (~8B), Mistral Latest (~7B), and Llama 3.1 8B (~8B). Setting up Ollama requires installing it from https://ollama.com and pulling desired models using commands like `ollama pull gemma3:4b`. To allow the extension to reach Ollama, users must set the OLLAMA_ORIGINS environment variable to include their extension ID before starting the Ollama service, either via command line or (for system-installed instances on Linux) by editing the systemd override file.

Apogee goes beyond simple truncation-based summarization. It embeds page content locally using a small on-device retrieval model and answers questions by ranking passages by relevance to the query. This means asking about details buried deep in a long article or PDF still works, rather than limiting results to whatever fits in the opening truncated slice. A "Highlight-in-page" feature (available on Chromium browsers) allows users to verify summaries by clicking any bullet point and seeing the original passage highlighted in the live page, enabling easy fact-checking without re-reading the entire article.

Installation is straightforward. Chromium users (Chrome, Edge, Brave, Opera, Vivaldi, Arc, Dia) download the packaged extension ZIP from Releases, extract it, open their browser's extensions page (chrome://extensions, edge://extensions, etc.), enable Developer mode, and click "Load unpacked" to select the extracted folder. Firefox users can install directly from Mozilla Add-ons or manually via Releases. For developers, the extension can be built from source with `npm install && npm run build`, which produces both dist/chrome and dist/firefox folders. No backend installation, terminal commands, or API key setup is required for in-browser mode; users simply click the Apogee icon, select "Summarize this page," and the model downloads and caches on first use. Keyboard shortcuts (default Alt+Shift+U, remappable) and right-click context menu options provide faster access without opening the popup.

Apogee emerged directly from the failure of Mozilla's Orbit project, which attempted browser-based page summarization but relied on centralized API servers and cached results server-side—a model Mozilla ultimately abandoned. The creator identified Orbit's core architectural flaw: cloud dependency makes a tool vulnerable to discontinuation and creates privacy risks, since user data and generated summaries passed through external endpoints. By reversing that architecture, Apogee eliminates the single point of failure—because it has no servers or cloud infrastructure to maintain, it cannot be "sunset" by corporate decision.

The technical foundation relies on two complementary execution paths. Chromium browsers (Chrome, Edge, Brave, and others) run quantized models via WebLLM, which leverages WebGPU to offload computation to the GPU. Firefox takes a different route through Transformers.js, executing smaller ONNX models on the CPU via WebAssembly, since Firefox's browser extension architecture lacks the offscreen document API that WebGPU requires. Both paths download model weights once (typically 270 MB to 2.2 GB) and cache them locally, allowing completely offline operation after the initial setup. For users with demanding workloads or older hardware, the Local Ollama mode bridges to larger models by communicating over localhost, preserving the privacy guarantee while gaining raw capability.

The extension also implements on-device retrieval—embedding the page content using a small local model and answering questions by ranking passages by relevance, rather than blindly truncating long content. This means users can ask about details buried deep in articles or PDFs and get accurate answers, not just summaries of the opening section. Highlight-in-page verification (Chromium only) allows spot-checking claims by clicking a bullet and seeing the passage from the original page.

No comments yet. Be the first to share your thoughts!