Apogee: A Local-First, Private AI Browser Assistant

Apogee is a local-first AI browser assistant designed for summarizing articles, videos, emails, and documents while maintaining total user privacy. Unlike cloud-based AI extensions, Apogee performs tokenization, inference, retrieval, and caching directly on the user's device, eliminating the need for API keys, subscriptions, or cloud data transmission.

Local Inference Architecture

Apogee leverages browser-native technologies to run LLMs locally, ensuring that page content and summaries never leave the device. It supports multiple execution backends depending on the hardware and browser:

  • WebGPU: Used by Chrome, Edge, and other Chromium-based browsers to utilize the GPU for faster inference.
  • WebAssembly (WASM): The default for Firefox and an opt-in fallback for Chromium browsers, allowing the tool to function on machines without a dedicated GPU.
  • Local Server Integration: Power users can connect Apogee to a local Ollama instance or a llama-server (via llama.cpp) over a local loopback connection (127.0.0.1).

Core Features and Capabilities

Apogee provides a suite of tools for content distillation and interaction across various media types:

Web Content and Social Media

Apogee uses Readability and site-specific extractors to clean text from Wikipedia, GitHub, Reddit, Hacker News, Bluesky, Mastodon, Lemmy, Discourse, Stack Overflow, and arXiv. It can summarize full pages or specific text selections (minimum 20 characters) via the browser context menu.

Multimedia and Documents

  • Video Summarization: Supports YouTube and Bilibili, providing interactive timestamped timelines that allow users to jump to key moments in the video.
  • Document Support: Users can drag and drop PDF, DOCX, TXT, Markdown, JSON, or HTML files into the extension for local processing.
  • Social Threads: Parses discussion sites into structured Markdown, preserving author identity, scores, and reply order.

Advanced Interaction

  • Smart Retrieval Q&A: Employs local passage matching to allow users to ask questions about long documents without losing context.
  • Grounding: On Chromium browsers, users can click summary bullets to automatically scroll the page to the corresponding source passage.
  • Customization: Supports custom standing instructions (system prompts) and per-page focus keywords to steer the AI's emphasis toward specific themes.
  • Translation: Offers multi-language translation into 32 target languages using the Helsinki-NLP Opus-MT engine or direct LLM translation.

Privacy and Security Model

Apogee is designed to prevent data leaks by restricting all processing to the local environment. Key security measures include:

  • Zero Cloud Transmission: No page contents, transcripts, or summaries are uploaded to cloud APIs.
  • Network Restrictions: Connections to Ollama and llama.cpp are strictly limited to http://127.0.0.1 to prevent external data transmission.
  • Sensitive Site Exclusions: To prevent accidental caching of private data, Apogee excludes sensitive domains such as Gmail, Outlook, Proton Mail, WhatsApp Web, Telegram Web, Slack, Discord, and Microsoft Teams from disk caching.
  • Anonymized Lookups: YouTube sponsor lookups use k-anonymity hash prefixes and can be disabled entirely in settings for a fully local experience.

Comparison with Mozilla Orbit

Apogee was inspired by Mozilla's discontinued Orbit project. While Orbit attempted browser-based summarization, it relied on central API servers and stored summaries remotely. Apogee diverges from this model by implementing a local-first architecture where inference and storage happen entirely on the user's device, removing the dependency on remote servers and the requirement for an internet connection after the initial model weights are cached.

Sources

Related

  • Dispatch
    OpenBrief: A Local-First Approach to Video Summarization and Knowledge Management

    Explore OpenBrief, an open-source desktop application that transforms video and audio into grounded summaries and searchable briefings using local-first AI.

  • Dispatch
    DocuBrowse 0.9.1: Local AI-Powered Document Search Engine

    DocuBrowse 0.9.1 is a local, FOSS document search engine that combines keyword and semantic search with AI-generated synopses, running entirely on-device via Ollama.

  • Project
    theJayTea/WritingTools

    Writing Tools is a free, open‑source desktop app for Windows, Linux and macOS that lets you invoke an LLM anywhere on your computer with a hot‑key. It can proofread, rewrite, change tone, translate, summarise web pages or YouTube transcripts, and even chat with the model. You can use cloud APIs (Gemini, OpenAI, Anthropic, etc.) or run local models via Ollama, llama.cpp or Apple‑silicon MLX. The app is privacy‑focused (no telemetry, local storage of API keys) and licensed under GPL‑v3.

  • Project
    linuxrebel/DocuBrowser

    DocuBrowse is a cross‑platform desktop/web app that indexes a wide range of document types and lets you search them by keywords and by AI‑generated semantic similarity, all using locally‑run Ollama models. It offers deep‑link passage navigation, on‑demand AI summaries, PII filtering, and runs entirely offline.

  • Project
    topoteretes/cognee

    Cognee is a free, open‑source platform that gives AI agents a persistent, searchable knowledge‑graph memory. It ingests text, code, and other artifacts, stores them locally (or in Postgres), and lets agents retrieve facts or generate answers, optionally using an external LLM. It provides a Python API, CLI, REST, and plugins for popular agent frameworks, and can be run entirely on‑premise with small CPU‑only models.