Lanbitou: A uTools Note Plugin That Automates Tagging and Keeps Notes Local
The flat-tag, AI-decides-organization model removes the friction that usually kills personal note systems, and local storage keeps the data portable. The cheap per-tag inference cost and the abandoned Cursor/Codex workflow both set realistic expectations for AI-assisted personal tools.
Lanbitou, an open-source uTools plugin, approaches note-taking from the fragment up. Opening the plugin focuses the cursor in a single input box; there is no title field, no formatting toolbar, just a place to drop a thought. Those fragments live under a flat set of tags, and the plugin deliberately supports no second-level nesting. When a note has no tag, an LLM assigns it to an existing tag or creates a new one, and the response format is checked and retried before it is accepted.
The same fragments feed long-form writing: while composing a document, you can pull up related fragments that AI judges semantically close, then have a template organize them. Todos are extracted from fragments by AI, linked to tags, and can be turned into daily or weekly reports from custom templates. All data stays in local files, with export and import for moving between machines. CNY 10 of DeepSeek credit lasted about six months of personal use, and the plugin accepts any user-provided API key.
Lanbitou runs as a uTools plugin to get quick launch and search on Windows and macOS. A more ambitious workflow that would turn meeting notes into delegated Cursor/Codex tasks through Feishu was abandoned after the Cursor SDK integration worked; the maintenance cost was judged too high.
Flat tags are the most opinionated decision here: organization happens after capture, and the LLM absorbs the cost of filing. That treats note-system abandonment as a design flaw rather than a user discipline problem.
CNY 10 lasting six months for AI tagging is a concrete sign of how cheap small-model calls are when the task is classification with a JSON schema.
The abandoned Cursor/Codex delegation flow shows the bottleneck for AI note tools is not invoking an agent SDK; it is the surrounding workflow of scheduling, error handling, and acceptance that eats the effort.