What it does
notabene turns a repo's documentation into a navigable site where people can select text on a rendered page, or comment on a whole page, diagram or image. The project treats each anchored comment as an instruction an AI agent can act on without extra quoting, instead of squeezing all feedback into one prompt. Comments and a change journal are stored as versioned JSON inside the repository, so there is no SaaS and no database. After you tell your agent to address the doc comments, it reads the comment store, edits the docs, marks each comment resolved and appends a journal entry covering what changed, why and which comments prompted it. An approve mode lets you check each proposed edit against its real git diff. Setup is through npm: install the package, run init to write the config and store, and run dev to open the site locally. It ships as a Claude Code plugin and skill, and the review protocol is file-based so any agent can follow it. The site renders Markdown and MDX, and lists search, Mermaid diagrams, multi-language docs and PDF export, plus a read-only public static build with llms.txt and Markdown copies of pages. Editing from the page is dev-only. The project says its write API is bound to loopback and refuses to write a file that git is not tracking.
Capabilities
- Anchored comments on text, pages, diagrams and images, with threads and resolve
- Agent review loop that edits docs, resolves comments and journals each change
- Approve mode to check each edit against its git diff
- Comments and journal stored as versioned JSON in the repo
- Markdown and MDX doc site with search, Mermaid diagrams, multi-language support and PDF export
- Read-only public static build that includes llms.txt
When to use it
- Reviewing project docs with teammates and letting an agent apply the feedback
- Keeping a trail of why documentation changed
- Publishing a documentation site straight from a repository
- Giving an agent located feedback rather than one long prompt
