I have kept my notes in Logseq for years. Daily journal, nested bullets, [[links]] everywhere, a pile of TODOs that I cycle with Cmd+Enter. I like the shape a lot. What I stopped liking is everything around it.

Why

Three things kept bugging me.

Sync. I write on a laptop and a phone. Logseq's syncs is paid, and non-existent on iOS on the DB version...

Agents. Half my day now runs through Claude Code agents (see talon). I want them to read my notes, find the page about a project, and add a bullet under the right parent. With a folder of markdown files the agent ends up rewriting whole files and hoping nothing else touched them in the meantime. That works until it doesn't, and with sync in the mix it doesn't pretty fast.

Hackability. Logseq is a big ClojureScript codebase. Every time I wanted a small change I gave up before I found the right file.

So I wrote my own. Well, me and Claude Code did: about 900 commits during two weekends, with me deciding scope, using it daily and filing bugs, and agents doing most of the typing. It's MIT licensed and lives at github.com/hnykda/nooklet, with docs at nooklet.danielalder.cz.

What it is

nooklet is an outliner for one person with several devices. If you have used Logseq, Roam or Obsidian's outliner, you already know how to use it: nested bullets, [[page]] and #tag links, linked and unlinked references, namespaces, properties, a journal stream with a calendar, and tasks with TODO/DOING/DONE markers, scheduled and deadline dates. It imports a Logseq file graph.

The interesting parts are underneath.

Local-first. Every device (desktop app, browser, iPhone) holds the whole graph in SQLite and works offline. Edits go into an op log with hybrid logical clocks. When a device reconnects, it trades ops with a self-hosted server and they converge. The server is optional: one laptop with no server is a perfectly normal way to run it.

Agents are first-class users. Every operation the app performs is defined once, and that one definition becomes an HTTP endpoint, an OpenAPI entry, an MCP tool and a method on the typed client. So Claude Code gets exactly what the UI gets: search (keyword, semantic or hybrid), read a page with stable block ids, append markdown under a block, replace an exact substring inside one bullet, run an atomic batch and undo it. Each agent has its own token, and its label shows up in page history as the author of its edits. Hooking it up is one line:

claude mcp add --transport http nooklet https://<your-server>/g/default/mcp \
  --header "Authorization: Bearer <token>"

A markdown mirror you can walk away with. The server writes every page to a plain markdown file with a stable ^id on each block. Logseq and Obsidian can open the folder, and grep and git work on it. SQLite holds the truth, the files are a copy, and if nooklet dies tomorrow your notes are still sitting there as text.

Everything optional is off. Sync needs a device you pair. AI access needs a token you mint. Semantic search needs an embedding model you run (any Ollama model). Skip all three and you have a local outliner with a markdown mirror.

Plugins. Features can ship as TypeScript plugins on both the server and the client. Mermaid diagrams, word count and a daily summary are built in as plugins, mostly to keep the plugin API honest.

What it is not

No flashcards, no whiteboards, no kanban, no real-time co-editing, no multi-user accounts, no hosted service. The scope stays at outliner, tasks, links, search, sync and an API agents can use. I would rather have a small thing that works than a big thing that almost works.

State of things

I use it daily on my Mac and my phone, and I run the server on my home Kubernetes cluster, reachable only over Tailscale. The everyday stuff works: editing, nesting, links and tags, journals, tasks, images, search, sync both ways. It is not bug-free, the phone UI still has rough edges, and you build the desktop and iOS apps from source for now. Expect breaking changes before 1.0.

If you live in Logseq and any of the above itches the same way it itched me, give it a go:

git clone https://github.com/hnykda/nooklet.git
cd nooklet
pnpm install
pnpm --filter @nooklet/web build
pnpm nooklet serve

Then open http://127.0.0.1:6100 and import your graph with pnpm nooklet import /path/to/graph. Issues and PRs are welcome on GitHub, and so is a well-argued "you should support X", even if the answer is "no, because".