Ambient Context

Why it's ranked here

This is a single developer's project that treats a sensitive capability, reading everything on your screen, with real discipline. The privacy documentation separates what is structurally guaranteed from what is a heuristic, and states its own known gaps instead of hiding them. The source carries unit tests for the parts most likely to break silently: protocol dispatch and the window-snapshot parser. It is MIT licensed, needs no account, and sends nothing anywhere except an optional update check. The cost is dependency: it only runs on Apple Silicon Macs, and half its value, the knowledge base and daily notes, needs an external agent CLI you must already own and configure.

What's good

The redaction model works in layers. Sensitive input fields are skipped before their text is read, whole snapshots are dropped for password managers and private browsing windows, and pattern matching scrubs credential-shaped strings afterward. Screen text is deduplicated across the day, so a page you keep revisiting is recorded once. Captured blocks keep the actual file path or web address behind them when the source app exposes one, so a reader can open the original document instead of trusting a scraped fragment. Read access keeps working even when the app is closed, while anything that changes settings or capture state gets logged with the connecting client's name attached.

Tradeoffs

Capture quality is uneven by design. Chromium and Electron based apps, browsers, chat clients and editors among them, build their accessibility structure lazily, so the first seconds after switching to one are thin and only fill in on later passes. GPU rendered terminal emulators expose little or no readable text at all, so a terminal heavy workflow will show real gaps in the record. The documentation is explicit that redaction is a heuristic, not a guarantee: an ordinary, non private browser tab showing banking, health or mail content is captured like any other page, and an app that mislabels a secure field defeats the strongest filter available.

How to use it well

This suits someone who runs an agent CLI day to day and wants it working from what actually happened, not a hurried recap. Point the capture folder outside cloud synced directories to keep the record on one machine, grant the permission it asks for on first launch, and set a processing time so the daily page and summary are ready each morning. Install the bundled skills so the assistant knows when to pull a summary rather than read everything. It is a poor fit if your day lives mostly in a GPU rendered terminal, or you are not ready to treat a folder of your own screen text as sensitive.

Observed

License

MIT (bundled fonts under a separate SIL Open Font License)

Platform

macOS 14 or later, Apple Silicon only

Packaging

Native desktop app built with Tauri: a Rust core and a React and TypeScript frontend, built via Node and Cargo

Interfaces

Menu bar GUI and an MCP server reachable over stdio, from the same compiled binary

Network surface

No account and no telemetry; the only outbound call is an optional GitHub Releases update check

Test coverage

Unit tests are present in the Rust source for the MCP protocol dispatch and the window-snapshot parser