Glonce

Ambient Context turns your screen into a markdown memory for LLMs

toolsThe Glonce editorsHacker News🔥61Glonce rating61Impact45Novelty60Relevance85Surprise35Players15Credibility75
Ambient Context turns your screen into a markdown memory for LLMs

A developer using the GitHub handle dragthelake has posted a Show HN for Ambient Context, a macOS menu bar app that keeps a written record of what the user works on so their own LLM can read it back. While the eye icon in the menu bar is open, the app reads the text of whichever window is currently focused, using the macOS accessibility tree, every few seconds, and appends what it finds to a plain markdown file: one file per day, saved in a folder the user chooses. Pointing Claude Code or any other agent at that folder lets it answer questions such as what the user worked on on a given day, build up memory about ongoing projects, or draft a standup update. The app captures no screenshots and no video; it reads only text, through the accessibility API. It makes no network calls in this build at all, has no account, server, telemetry or bundled model, though the eventual signed release will add a single update check against GitHub. Before anything is written to disk, password managers and private browsing windows are excluded entirely, password fields are skipped at the source, and credentials, API keys and card-shaped numbers are scrubbed. Lines are deduplicated across the day so the file stays small enough to hand to an LLM whole, interface junk is filtered out, and each captured block records the document path or URL it came from so an agent can open the original rather than trust a fragment. The capture folder documents its own markdown format in an included AGENTS.md file, and the folder is itself excluded from capture so the app cannot record its own output. The app reads only the focused window, never background windows, other displays, minimised windows, or anything while the screen is locked. It requires macOS 14 or later on Apple Silicon. There is no notarised download yet, Apple Developer enrolment is described as in progress, so for now a user has to build it themselves from source with Node, Rust and Xcode Command Line Tools, a process the README says takes about two minutes: clone the repository, run npm install, then npm run tauri build, and drag the resulting app into Applications. The default save location is outside the user's Documents folder specifically so iCloud does not sync the captured record off the machine. The README also flags known gaps in the underlying mechanism: Chromium and Electron apps such as Chrome, Slack, VS Code, Obsidian and Figma only build their accessibility tree on demand, so the first seconds of capture in those apps come back thin and fill in on later passes, and Chrome can show a minor glitch in its window-resize animation while capture is enabled. GPU-rendered terminals like Kitty and Alacritty expose little or no text to the accessibility API, while Terminal.app and iTerm2 work normally. The project asks users to report which apps come back rich, partial or empty, with a template provided in the repository's docs.

Key facts

  • Ambient Context is a macOS menu bar app that reads the focused window's text via the accessibility tree every few seconds and appends it to a plain markdown file, one per day.
  • The captured folder is meant to be pointed at Claude Code or another LLM agent, which can then answer what the user worked on, build project memory, or write a standup.
  • It captures text only, never screenshots or video, makes no network calls in the current build, and has no account, server, telemetry or bundled model.
  • Password managers and private browsing are excluded entirely, and credentials, API keys and card-shaped numbers are scrubbed before anything is written to disk.
  • It requires macOS 14+ on Apple Silicon and currently has no notarised download, so users build it from source themselves, a process the README estimates at about two minutes.

Why it matters

Ambient Context targets a specific gap in the fast-growing space of LLM agent memory: giving an assistant like Claude Code durable, low-effort context about what its user has actually been doing, without the privacy and storage overhead of screenshot-based recall tools. By writing plain, human-readable markdown rather than images or a proprietary database, it keeps the record inspectable, editable and portable, which fits a broader trend of coding agents wanting structured, greppable context rather than opaque logs.

Who it affects

The primary audience is developers and other power users on Apple Silicon Macs who already work with LLM coding agents such as Claude Code and want those agents to have memory of the user's own screen activity across a workday. It also affects anyone weighing screen-recording style productivity or memory tools who is specifically concerned about screenshots and video leaving a trace, since this project's whole design premise is to avoid both.

How to use it

The app is free and open source on GitHub, with no license or pricing terms stated in the material available. It is not yet notarised or signed, so installation currently means cloning the repository and building it locally with Node, Rust and Xcode Command Line Tools, a process the README says takes about two minutes, then dragging the built app into Applications. On first run it asks for macOS Accessibility permission, without which it cannot read window text, and lets the user pick a save folder, defaulting to a location outside Documents so iCloud will not sync the captured record. Clicking the menu bar eye toggles recording on and off.

How solid is it

The material available is the project's own README and a Show HN discussion with 21 points and 7 comments at the time of writing, so the claims here are the developer's self-description rather than independent verification. The README documents specific, checkable behaviour, including the exact build commands, a sample captured markdown block, and named limitations for particular apps and terminals, which suggests a working prototype rather than a concept, but there is no notarised release, no stated user base, and no license terms given.

Risks and caveats

The project is early and unsigned, requiring users to build it from source themselves, which is a meaningfully higher bar than installing a signed app and limits how many people will actually run it as described. The underlying accessibility-tree approach has acknowledged gaps: Chromium and Electron apps take a few seconds to populate fully, and GPU-rendered terminals such as Kitty and Alacritty expose little or no text at all, so capture quality varies by application. No developer or company name, license, or pricing terms are given beyond the GitHub account the code is hosted under, and no notarised build has a stated release timeline.

“The app reads only the focused window: never background windows, other displays or minimised windows, and never while the screen is locked.”

— project README