OpenKnowledge for Windows: Local-First AI Markdown Editor for Obsidian Vaults

What OpenKnowledge is​

OpenKnowledge comes from Inkeep, an AI agent company. The GitHub project describes it as a Markdown editor with integrations for Claude, Codex, and other harnesses, for knowledge bases, LLM wikis, specs and notes. The v2.0 public release was on June 3, according to the project's website, which matches the review's note that it showed up in June. Other outlets give slightly different launch dates, so treat the exact day loosely.

The licensing and maturity details matter for anyone planning a long-term notes archive:

• It is free and open source under GPL-3.0.
• Coverage of the repository at v0.18.0 shows 296 stars, 613 commits and 170 releases. The project is young and moves fast. The contributor guidance also treats it as pre-1.0.
• There are desktop apps for macOS, Windows and Linux, plus a web UI.

One caution. Some early third-party coverage described it as macOS-first, and one outlet said the desktop app ran only on Apple Silicon. Current project documentation lists Windows and Linux desktop builds, and the reviewer ran it on their PC. Still, the Windows build is newer than the Mac one, so expect rough edges.

Files, not a database​

This is the main reason an Obsidian user might trust it. The project's docs say there is no database dependency at all. The knowledge base is plain Markdown in your own project directory, and the file system is the only persistence layer. Git handles version history.

The reviewer's experience fits that design:

• Opening a vault: They pointed it at an existing Obsidian vault. Notes, non-Markdown files, wikilinks and YAML frontmatter all appeared intact.
• Obsidian settings: OpenKnowledge ignores Obsidian's settings folder and keeps its own configuration in a hidden .ok folder.
• What doesn't carry over: Anything plugin-specific that you built in Obsidian.
• Exports: There is no export button, because there is nothing to export. Notes are already files in the folder.

That is the reviewer's own test on one vault. It isn't a compatibility guarantee. If your vault leans on plugins, dataview-style queries or unusual link conventions, test on a copy first.

The docs add a warning for Windows users with older notes. OpenKnowledge's link resolution has specific rules about case-sensitivity and about wikilinks that include .md suffixes. On Windows and macOS, a Markdown image or asset link whose folder spelling differs in case from the folder on disk can show as broken after an upgrade. After moving a vault in, spot-check your links.

The editor​

The editing experience is closer to Google Docs or Notion than to Obsidian's live preview. Formatting renders as you type, and you can flip a note to raw Markdown and back. Per the project's documentation, the editor also offers:

• a frontmatter panel for properties
• an outline and lists of incoming and outgoing links
• a graph view
• a timeline of recent edits with per-burst diffs and selective rollback
• rendered extras such as Mermaid diagrams, LaTeX, callouts, collapsible sections and interactive HTML

The reviewer also liked per-folder templates and the slash-command menu.

How the AI part works​

The agent integration is where it differs from a chat box bolted onto a notes app. OpenKnowledge exposes its engine over MCP, the Model Context Protocol. Native MCP lets you plug the knowledge base into Claude, Cursor, Codex and other agents. The core-concepts page says this keeps frontmatter consistent, references intact and the link graph healthy as agents edit.

The reviewer reports that the setup dialog found Claude, Codex and OpenCode already installed on their PC. It used those installs instead of billing separately, so costs fall on subscriptions you already have. The app also supports local models through LM Studio. The reviewer, citing the docs, warns that smaller models tend to fumble tool calls, so results depend on your hardware.

In the reviewer's competitor-research test, the agent did the following:

1. Built a folder structure with a template in each folder.
2. Wrote an overview page with a diagram.
3. Added short folder descriptions aimed mainly at the AI.
4. Ingested sources by URL, saving the full page with the URL and capture date.
5. Wrote research notes that cite those saved sources and stay drafts until the user signs off.

The project describes this approach as "closed-loop grounding". Every claim should trace to a source stored inside the knowledge base. The docs frame this as a supported workflow pattern, not a requirement. So the source-citation discipline the reviewer saw is a feature of how the knowledge base was set up, not a guarantee the app enforces everywhere.

For oversight, an agent icon opens a list of the files it touched, with diffs. You can undo changes, and deletes and moves ask for permission. A diff shows what changed. It doesn't show whether the new content is correct, so you still have to read what the agent wrote. The reviewer had only used it for about a week and called their knowledge base sparse.

Setting it up on Windows​

The project's quickstart gives two routes.

Desktop app (recommended):

1. Download the Windows installer and launch the app.
2. Confirm Git is installed by running git --version in a terminal. The app uses Git for its timeline and recovery features.
3. On first launch, a setup dialog offers to connect your AI tools and to add the ok command to your terminal. Every option starts unchecked, so tick what you want and click Finish setup. You can reopen it later from the File menu.
4. Create a new project or open an existing folder, then use the Ask button to hand a prompt to an agent.

The docs add that on Windows the installer puts the bundled CLI on your user PATH itself.

CLI and local web app:

• This route needs Node.js 24 or newer and Git.
• Run npm install -g @inkeep/open-knowledge.
• In your project folder, run ok init, then ok start to serve the editor in your browser.

If an agent asks you to approve a new MCP server, approve the one named open-knowledge. Cursor doesn't ask.

What "local-first" does and doesn't mean​

The title's "trust" is the claim that needs the most care. OpenKnowledge's own documentation says nothing about your content leaves your machine by default. It then lists exceptions:

• Opt-in or on-demand: semantic search, diagnostic or bug-report bundles, and GitHub sync can send data off the device.
• Automatic, with no document content: the desktop app checks for updates, sending its version and channel. On first launch it opens your browser once to check for a pending share link.
• Crash dumps: a crash dump can contain document content and can't be redacted. The docs say it's attached only if you check the box, and you review the report before sending.

The bigger caveat sits outside the app. Storing notes locally says nothing about where your prompts and note excerpts go once an agent such as Claude or Codex reads them. That depends on the agent, model, account and settings you chose. A fully offline setup would need a local model, with the tool-calling limits described above. I found no independent security audit of the app. Nor did I find anything validating the accuracy of agent-written research.

Who should try it​

My read, based on the evidence rather than any testing of my own:

• A good fit: Obsidian users who already use Claude, Codex or similar tools and want them working on notes without moving to a cloud service.
• A cautious fit: anyone with a plugin-heavy vault. Test on a copy.
• A harder sell: people who want a finished 1.0 product with long-term stability guarantees.

The reviewer's own conclusion is modest. It won't replace Obsidian, but because it uses the same folder and the same Markdown, it works alongside it. Since it only reads and writes ordinary files, trying it costs little. Open a copy of your vault, check that Git history works, review the privacy settings, and decide from there.