Sentient-OS-Labs/sentient-os

An on-device LLM understands your entire life, then proactively offers to get your work done through computer use.

What it solves

Sentient OS addresses the privacy and context gaps in current AI assistants. Most AI tools lack deep personal context because users cannot upload their entire digital lives to a cloud, and they are reactive rather than proactive. Sentient OS creates a local, private knowledge base of a user's life to enable an AI that understands the user's specific context and can proactively suggest tasks or execute them via computer use.

How it works

The system uses a hybrid compute architecture where approximately 90% of the work happens on-device and 10% uses a frontier model via the user's own ChatGPT subscription.

  1. Local Processing: Every night at 3 AM, the system wakes the Mac and uses a local model (Gemma 4 E4B) to read files, screenshots, WhatsApp, iMessage, and Apple Notes. It filters out junk and sensitive PII, creating clean summaries.
  2. Knowledge Base: These summaries are distilled by a frontier model into a folder of plain markdown files (Obsidian-style) stored locally on the Mac.
  3. Proactive Intelligence: The frontier model analyzes the knowledge base to prepare "morning cards"—drafted replies or task reminders—that the user can execute with one click.
  4. Sidekick: A voice/text interface (accessible via the Mac notch or a keyboard shortcut) that uses "computer use" capabilities to perform tasks within the user's own apps and browser.
  5. Optional Cloud Mirror: Users can optionally sync their knowledge base to a cloud MCP server using zero-access AES-256-GCM encryption, allowing other AIs like Claude or ChatGPT to access the personal context.

Who it’s for

Mac users with Apple Silicon (M1 or newer) running macOS 15+ who want a highly personalized, proactive AI assistant while maintaining strict data privacy.

Highlights

  • Privacy-First Architecture: Raw personal data never leaves the Mac; only PII-stripped summaries are sent to the cloud.
  • On-Device Inference: Optimized for Apple Silicon, running a custom LiteRT-LM fork of Gemma 4 E4B on devices with as little as 8GB RAM.
  • Proactive Workflow: Instead of waiting for prompts, it prepares task cards overnight based on the user's personal context.
  • Computer Use: Sidekick can interact with local applications and browsers to complete tasks autonomously.
  • Transparent Memory: The knowledge base is stored as editable markdown files, not a black-box database.