# Ogle

Ogle is a private macOS state layer for MCP-capable AI assistants. It samples local attention, app activity, browser activity, system pressure, and optional camera or screenshot-derived signals, then stores compact derived rows and searchable visual captions in local SQLite.

Use Ogle when an assistant should adapt pace, detail, option count, and next-step size around a user's current state estimate. Treat all state fields as estimates. Do not call them facts about emotion or health.

## Repository

- Source: https://github.com/sliday/ogle
- Landing page: https://ogle.dev/
- macOS entry point: `Sources/Ogle/main.m`
- MCP server: `mcp/ogle_mcp.py`
- Browser native host: `native-host/ogle_browser_host.py`
- Browser extension: `browser-extension/chrome`

## Requirements

- macOS 13 or newer
- Xcode Command Line Tools
- Ollama serving `gemma4:latest` and `embeddinggemma:latest`
- Optional Chrome or Comet extension for reliable website rows
- Codex, Claude Code, Gemini CLI, or another MCP-capable assistant

## Build

```sh
./scripts/build.sh
open "build/Ogle.app"
```

## MCP Tools

- `ogle_get_state`
- `ogle_activity_stats`
- `ogle_codex_context`
- `ogle_taxonomy`
- `ogle_brain_context`
- `ogle_state_layer`
- `ogle_state_insights`
- `ogle_visual_memory_search`

## Privacy Boundary

MCP responses should contain derived summaries, scores, facts, aggregates, visual-memory captions, and recent context. They must not expose raw screenshots, webcam frames, audio, key values, prompts, JPEG blobs, or full URLs.

## Assistant Behavior

Read `ogle_codex_context` for Codex-specific pacing. Read `ogle_state_layer` for the general state contract. Use the `STATE:` label in user-visible summaries. Reduce branching when focus is low, slow down when fatigue is high, and keep the signal as a weak hint when confidence is low.