An 'AI-first, open source' notes app that can't talk to a local LLM

Inkeep shipped OpenKnowledge as a GPL-3 Obsidian rival whose whole pitch is AI collaboration. The catch the thread found fast: it only drives Claude, Codex, and Cursor desktop apps, with no way to plug in a model you run yourself.

In brief

OpenKnowledge is a local-first markdown editor that auto-syncs a vault over Git and hands editing duties to AI agents. The signature README gif shows an agent drafting a launch recap while you watch.

Then the thread asked the obvious question: it's labeled open source and 'AI-first', so why can it only call Claude, Codex, and Cursor, and not a local Gemma or any OpenAI-compatible endpoint? One veteran wanted 'consequences' for shipping that combination under the open-source banner.

A GPL-3 notes app whose killer feature is letting an agent type for you

OpenKnowledge editor with an AI agent drafting a document
OpenKnowledge editor with an AI agent drafting a document

OpenKnowledge — ok on the command line — is Inkeep's bid to be the open, AI-native answer to Obsidian and Notion. Under the hood it is plain markdown in folders, edited through a WYSIWYG surface built on tiptap with CodeMirror underneath for raw code. A managed Git layer syncs a vault to GitHub and auto-commits, so a non-technical teammate gets versioning without ever touching git. The code is TypeScript, the license is GPL-3.0-or-later, and the repo is moving fast: 175 releases against 638 commits.

The headline word is AI-first. In practice that means the editor knows how to launch and coordinate Claude, Codex, and Cursor desktop apps to read and write your notes, plus an MCP (Model Context Protocol) server, a skills framework, and agentic search over the vault. The signature gif at the top of the README shows an agent composing a launch-recap doc in the editor while the human looks on.

One sharp detail surfaced from a tester: the desktop installer quietly adds ok to your shell PATH. As one user with a 3,200-karma account put it, the toast read "Added ok to your PATH — managed block in ~/.zshrc" and "Took a while to see that 'ok' is the name of your product."

The hole everyone poked at: no local model, on an 'open source' app

For a project that leads with open and AI, the thread converged on one absence. It can orchestrate three commercial coding assistants but offers no way to point it at a model you host.

"Fully local, but can't integrate with any local LLM?" — an HN user with \~6,700 karma, posting since 2014

The strongest version of the objection came from a long-tenured account, framing it as a naming problem rather than a feature gap:

"there genuinely ought to be consequences for using 'open source' in the context of something like this tied to proprietary AI services. Local models should be the first choice in that framing." — a commenter with nearly 8,000 karma

People described the exact rigs they wanted to wire in. One who runs LMStudio with Anthropic- and OpenAI-compatible endpoints reported Gemma running well for these tasks locally, and called the missing plug-in option "a deal breaker." Another self-described team lead hunting a Git-synced knowledge base liked almost everything but flagged the same gap: no way to connect any OpenAI-compatible provider.

'Isn't Obsidian already a pile of markdown?'

The second recurring jab questioned the premise that markdown notes need to be made AI-first at all.

"I don't understand how Obsidian, a collection of markdown files, isn't already AI friendly. It's hard for me to imagine a more AI-friendly but still usable way to organize your notes." — a user with a 10k-karma account linking a personal GitHub

The most detailed teardown came from someone who actually installed it and went looking for the value. Their verdict was that the new part is mostly a bundle of skills and MCP servers wrapped around a thinner Obsidian:

"Right now, it feels like a well dressed wrapper for pretty basic functionality… you want people technical enough to use LLMs and GitHub, but not so technical they can't customize a shared environment." — a tester who switched back to an Obsidian plugin after trying it

A blunter version: "So it's just a Electron editor + 'open in %agent' button… I don't see any reasons to use it instead of obsidian + my agents," wrote a commenter who already runs his own agents against Obsidian. Two separate notes pushed Inkeep away from Electron entirely toward a real native app, arguing the small affordances around selection and scrolling are where Electron knowledge tools always lose to the incumbents.

A name three different projects are now fighting over

There is also a branding collision the thread caught. Open Knowledge is not a free namespace. A former staffer at the Open Knowledge Foundation noted the okfn.org group has carried that name since 2004, building open-data licenses and the CKAN data-portal platform. Separately, another commenter asked whether this follows the Open Knowledge Format that Google proposed earlier the same month, or is just a collision — it is a collision.

The practical questions that will decide adoption barely got answered in the captured thread: whether an agent editing a file round-trips YAML frontmatter and nested code fences without mangling them, and whether there is a real migration path out of Obsidian or Notion, since switching cost is what keeps people locked in. For now the macOS app is native and everyone else runs the web app through the CLI on Node 24+. The Linux desktop notify-list is already filling up, which tells you where the demand sits: on the platform where the GPL-3 'open' app is hardest to actually run.

Get breakdowns like this every day

The top Hacker News stories, unpacked — by email, Telegram, or RSS.

Next story · 21 of 27 Read more
🌍Other

A Dubai brunch ended 24 dirhams short, and named the fallacy that caused it

Six friends sat down after yoga and spent the meal scanning a single QR code in turns, fielding a parking alarm mid-conversation, and untangling who had paid for the shared apple crumble. The bill ended 24 AED short.

★ 69 · 💬 100 · rozumem.xyz · 6 min