Apogee puts summaries and follow-up questions next to what you are reading, without requiring a hosted AI account. The project describes local extraction, inference, and caching for articles, discussion threads, captioned videos, and files supplied through its popup. It is useful if you want a browser reading assistant backed by models you control rather than another metered API.

This is an existing project with recent releases, not a brand-new launch. The September 0.2.2 release added llama.cpp integration, an optional Chrome side panel, and text/JSON summary exports. The latest 0.2.4 release, published October 5, fixes focus-keyword handling and cancels a running stream when its relay popup disconnects; downloadable Chrome ZIP and Firefox XPI artifacts are available.

  • Setup: Install through the Chrome Web Store or Firefox Add-ons links in the repository. Chrome/Edge use WebGPU—browser access to the GPU—by default; Firefox uses WebAssembly-based CPU inference. Model weights download before offline use.
  • Use your own model: In Settings, select Local Ollama and point it at http://127.0.0.1:11434, then choose an installed model. The llama.cpp option connects to a local llama-server at http://127.0.0.1:8080 through its OpenAI-compatible chat endpoint. No separate Apogee backend is required.
  • Cost: The extension is MIT-licensed, with no account, subscription, or paid API key required. Your costs are local compute, memory, storage, and the initial model download—not zero hardware requirements.
  • Privacy qualification: “Local” does not mean nothing is stored: summaries and extracted page text can be cached in extension storage. The privacy documentation also discloses model downloads and an optional SponsorBlock lookup enabled by default for YouTube. Turn that lookup off under Settings → Privacy if you do not want it.

The important limit: connections are restricted to loopback addresses, so you cannot simply enter the address of a GPU server elsewhere on your network. The llama.cpp guide suggests an SSH port forward for that arrangement. Safari is unsupported, and Firefox’s default inference path is CPU-based rather than WebGPU. These are documentation-based findings, not a hands-on quality or performance test.

Browser compatibility · Local model setup · Privacy and storage details