note.md
English · 简体中文 · notemd.net
Read what AI writes. Keep what you think. Keep what only you can write.
A markdown reader, editor, and bidirectional-linking notes tool designed for
people and AI agents to work in the same files. Desktop app for macOS 13+
(Apple Silicon and Intel), Windows 10/11 x64, and Linux beta. Your notes are a folder of plain
.md files on your disk — forever.
Download · Plugins · Full feature list
Linux beta packages provide .pkg.tar.zst for Omarchy / Arch and .deb for Debian / Ubuntu, each for x86_64 and aarch64; aarch64 has not been verified on a native desktop machine. Packages update through the package manager. Native plugins without a Linux build and app-internal agent panels are excluded; Wayland shortcuts use compositor bindings. See Linux installation, limitations and acceptance.
Download packages
| Platform | Architecture | Package |
|---|
| macOS 13+ | Apple Silicon | .dmg |
| macOS 13+ | Intel | .dmg |
| Windows 10/11 | x64 | .exe |
| Linux beta · Omarchy / Arch | x86_64 | .pkg.tar.zst |
| Linux beta · Omarchy / Arch | aarch64 | .pkg.tar.zst |
| Linux beta · Debian / Ubuntu | x86_64 (amd64) | .deb |
| Linux beta · Debian / Ubuntu | aarch64 (arm64) | .deb |
Host and plugin releases are independent. Check the changelog and each plugin's minimum host version before installing; a new Apple Silicon release does not imply an Intel, Windows or Linux update. Mobile builds are not published.
Windows and Linux links select the latest published version that actually has a package for that platform and architecture; it may differ from the newest overall release. Linux downloads ask for the package format when it is not specified. ARM never substitutes x86_64, and Arch packages never fall back to Debian packages. If none is available, the link opens all releases. Linux uses package-manager updates, outside the desktop updater manifest. The Omarchy repository entry has not been submitted, and Wayland/fcitx5/tray desktop acceptance is pending.
1. The best place to read what your agents wrote
Rich view and source view, one keystroke apart. Import Typora-compatible
themes. Mermaid code blocks render with Pele by default; Graphviz and KaTeX
are also supported and loaded on demand. See diagram compatibility.
The app uses the system WebView without bundling Chromium.
Highlight a claim, leave a question in the margin, fix the sentence right
where it's wrong.
Claude, Codex, OpenClaw each have their own chat window. None of them is a
place to read. This is.
2. Everything the last generation got right, built in
Local-first. Git sync. Outliner. [[wikilinks]] and backlinks. Wiki pages.
Daily notes. Full-vault search. JSON Canvas. A plugin system.
Roam Research and Obsidian figured these out. note.md ships them on files:
Roam Research Sync initializes from a full graph export and can then keep
changed pages in sync, while an Obsidian vault opens directly.
3. Agent-ready by design. Use the AI you already have.
Built-in agent workflows use the agents, AI subscriptions, API accounts, or
local runtimes you already have. note.md sells no tokens, adds no token markup,
and charges no separate per-token fee.
Model usage still follows the plan, limits, and billing of your chosen
provider—or uses your own local compute.
Your vault is designed to be the shared, version-controlled context that many
agents and many harnesses work in — Claude Cowork, Claude Code, Codex,
DeepSeek, ChatGPT Work, OpenClaw, Hermes — through public conventions (AGENTS.md,
block citations, sidecar .note.md). Memory follows the same contract:
agents can suggest what matters, but only you decide what becomes lasting
context. The more you leave in the vault, the better every agent knows you.
The vault holds more than what agents write. It also holds how they should
work: AGENTS.md, skills, schedules, MCP connections, memory. That playbook is
the judgment you left behind after every correction. Kept in the vault, it isn't
stuck in one harness's private folder.
Switch AI tools whenever you like. The asset stays yours, without adding a
second token meter from note.md.
Specifically tuned for building products with Claude Code: writing docs,
reading docs, and reviewing what the AI generated.
4. Memory discovered by agents. Confirmed by you.
What an agent most needs to know about you is exactly what no model can guess:
how you work, what you prefer, what you have decided, and where your boundaries
are. Asking you to maintain another profile only moves the work back to you.
Memory takes the opposite approach. An agent finds a small number of candidate
memories in the work and conversations you bring into your vault. You review
them one at a time: confirm, reject, mark important, or ignore. An inference
never becomes trusted memory on its own.
Search and Memory stay separate. Search helps an agent find anything you
wrote; Memory gives it the small set of claims you approved. They remain plain,
Git-tracked files that work across agents and models. Why we built personal AI
memory this way.
5. Whatever else you need, grow it yourself
Write a plugin. Wire an OpenClaw cron job. Hang skills off it.
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
