If you’ve ever set up a “second brain” for your AI agents, you know the drill: everything into Obsidian, slap on a community MCP plugin, and cross your fingers hoping Claude Code or Cursor reads your vault the way you expect. It works, but it’s duct tape.
Inkeep — a YC-backed startup building infrastructure for AI agents — just open-sourced a tool designed to remove that duct tape. OpenKnowledge hit Hacker News today and already has 81 points and 37 comments. The pitch, in the founder’s own words: “open source AI-first alternative to Obsidian/Notion”.
What it actually is
OpenKnowledge is a true-WYSIWYG markdown editor — local-first, free, and GPL-3.0 licensed — that installs as a native macOS app or as a cross-platform CLI. You init a project and it scaffolds the wiki structure, automatically connecting Claude Code, Cursor, and Codex:
npm install -g @inkeep/open-knowledge
cd your-project
ok init # scaffolds the project + connects Claude Code, Cursor, and Codex
ok start --open
That CLI/web path is what you’ll use on Linux, Windows, and Intel Mac. If you have Apple Silicon, you get the full native app via DMG — that asymmetry matters if your team isn’t 100% on Mac.
Why this isn’t “just another Obsidian plugin”
There’s a great exchange in the HN thread: someone basically asks, “isn’t a folder of markdown files already AI-friendly?”. Founder Nick’s answer is the whole thesis of the project: that Obsidian being closed-source means it can’t do what OpenKnowledge does — auto-install skills and an MCP server directly into Claude’s, Codex’s, and Cursor’s configuration, and open a live web viewer inside those agents’ embedded browsers, so you and the agent can edit the same document side by side in real time. It’s not a plugin taped onto a vault; it’s integrated into the agent-facing layer by design.
The interesting engineering problem
Two things stand out from the founder’s own technical breakdown, and they’re worth dwelling on because they explain why this required real work:
- ProseMirror (the editor’s rich-text engine) represents documents as an AST, which isn’t natively designed for byte-for-byte fidelity. So converting between the WYSIWYG view and raw markdown without losing or breaking formatting on each round-trip is genuinely hard.
- To solve it, OpenKnowledge runs a dual-observer CRDT (built on yjs) that keeps ProseMirror’s state and raw markdown state in sync, losslessly, in both directions.
That same CRDT — combined with git/GitHub running underneath — also powers undo/redo, version history, and “no-code” team sync without anyone touching a terminal. Your data lives in a regular git repo that you own; there’s no proprietary cloud sync layer holding it hostage.
The full OSS stack for anyone wanting to dig into the repo: Tiptap/ProseMirror, CodeMirror, yjs, Electron, Orama for search, and the remark/rehype/micromark/mdast toolchain for markdown processing.
What’s still unresolved
It’s a fresh OSS launch (the thread has been up for just hours as I write this), so there are a couple of things that are open questions rather than settled facts: how agentic search/RAG actually runs underneath — locally or via external model calls — isn’t documented yet, and the “all OSS for now but team sync is our business model” angle is worth watching closely, because GPL-3.0 could be friction for teams wanting to embed it in proprietary tooling.
What’s next
In the thread, Nick mentioned the team will prioritize integration guides with Hermes and OpenClaw next — both already covered on yoDEV: our piece on Hermes Agent and our article on OpenClaw.
Try it out
- Repo: GitHub - inkeep/open-knowledge: Beautiful, AI-native markdown editor and LLM Wiki · GitHub
- Site: openknowledge.ai
- Discussion: the HN thread is worth reading in full — it’s a good snapshot of how devs are already thinking about “AI second brain” architecture
Are you still using Obsidian + loose plugins for your knowledge base with agents, or have you already migrated to something native like this?