# OpenWiki turns a copy action into a wiki page you can decline

> OpenWiki is a Tauri and Rust desktop app for macOS and Windows that watches the clipboard, shows a ten second popup, and files whatever you keep into a local SQLite store. An AI layer then compiles that material into wiki pages, a knowledge graph and weekly attention reports, with Claude, OpenAI and Gemini as interchangeable providers.

**kdsz001/OpenWiki** — OpenWiki — Mac desktop AI knowledge management tool. Capture clipboard, build personal wiki, get AI insights.

- Repository: https://github.com/kdsz001/OpenWiki
- Website: https://openwiki.pages.dev
- Stars: 642 · Forks: 105
- Language: Rust
- License: MIT
- Published: 2026-10-21 · Updated: 2026-10-21 · Language: en
- Canonical page: https://hysenlabs.com/projects/kdsz001-openwiki

## A copy action opens a popup that closes itself after ten seconds

The capture loop is the whole product in four steps. You copy something, a popup appears on the desktop, you decide whether to keep it, and the AI organises what you kept into a knowledge base. The framing line in the README is blunt about where the agency lies: you decide what to keep, the AI makes sense of it.

Two details make that claim concrete. Only content you actively choose to keep gets saved, so there is no silent hoarding of every clipboard entry, and the popup auto-dismisses after ten seconds, so a copy made in passing does not queue up into a backlog you have to review later.

The capture accepts text, images and URLs, and it detects the source application automatically. For links it goes further and fetches the full article content, with WeChat and X/Twitter called out explicitly. That fetch is what turns a link into a body of text the model can summarise, rather than leaving you with a bare URL and a title.

There are shortcuts for the manual path as well: `⌘⇧C` on macOS or `Ctrl+Shift+C` on Windows opens the capture window on demand, and `⌘⇧Y` or `Ctrl+Shift+Y` brings back the main window, since closing the window only leaves the app running in the system tray.

## The AI layer turns captured material into wiki pages and a knowledge graph

Captures are not the end product. The AI automatically compiles what you keep into wiki pages organised around concepts, entities and topics, so a pile of clipped text becomes something with structure. On top of that sits a knowledge graph visualisation for seeing how ideas connect, and an Ask sidebar where questions about your own material are answered from your own content.

The tool also checks its own work. Orphaned pages, broken links and other structural issues are detected automatically, which is the kind of housekeeping a hand-kept notes folder never gets. In the interface this is a set of views rather than a document editor: filter captures by type and by time range, search globally across both the raw content and the knowledge base, and export everything to Markdown in one click.

The stack underneath is visible in the dependency list. The graph view is drawn with d3-force, animations come from framer-motion, the interface is React with Tailwind, and state is handled by zustand. Nothing here is unusual for a modern desktop front end, and the point of listing it is that the visualisation is a real feature backed by a real library rather than a screenshot in the README.

## Weekly reports grade your own reading habits across seven dimensions

The Insight Reports feature generates a one-click weekly AI summary of what you captured, then goes further with an attention analysis split into seven dimensions: At a Glance, Subconscious, Graveyard, Blind Spots, Hot Topics, Heatmap and Action Items.

The names describe the intent. Graveyard presumably covers the captures you saved and never returned to. Blind Spots and Subconscious are harder to read, but they sit next to Hot Topics and a heatmap of activity, so the report as a whole is trying to show the shape of your consumption rather than a summary of it.

The feedback loop is what makes it more than a digest. You can like or dismiss individual report items, and the AI learns your preferences from that, so the next report can follow what you found useful rather than repeating the same categories.

This is the part of OpenWiki that requires the most trust in the model, since it is the AI describing your own behaviour back to you. It is also the part most likely to be wrong in a way you will not notice, which is a reason to read the report rather than skim it.

## Three AI providers, two ways to authenticate, one local database

OpenWiki does not lock you to one vendor. Anthropic (Claude), OpenAI and Google Gemini are all supported, and for each one you can connect with either an API key or an OAuth login. You can also choose a different model per provider, so a cheap model can handle bulk classification while a stronger one handles the summaries.

The OAuth route needs identifiers that the app cannot ship for you. The .env.example file shows the three variables the app expects: GEMINI_CLIENT_ID and GEMINI_CLIENT_SECRET for Google Gemini OAuth, and OPENAI_OAUTH_CLIENT_ID for the OpenAI and Codex features. The file tells you to copy it to .env and fill in your own values.

Storage is local. Everything is kept in a local SQLite database, and the product description leads with privacy first as a result. The raw capture, the wiki pages, the graph and the reports all live on your machine rather than in a vendor's cloud, which is the reason the export to Markdown is worth knowing about: it is your escape hatch if you decide to move your material elsewhere.

The desktop behaviour matches the storage story. The app runs from the system tray, supports a dark, light or system theme, and exposes MCP protocol integration so it can connect to Claude Desktop.

## The macOS build is notarized, the Windows build is not

Getting OpenWiki onto a machine differs by platform, and the difference is not cosmetic. On macOS the app is signed and notarized by Apple, so you open the .dmg, drag OpenWiki into Applications and launch it without any security override. The platform split in the downloads covers Apple Silicon and Intel separately, with the archive names `OpenWiki_X.Y.Z_aarch64.dmg` and `OpenWiki_X.Y.Z_x64.dmg`.

On Windows the build is unsigned, so Microsoft Defender SmartScreen may warn on first launch. The guidance is to run `OpenWiki_X.Y.Z_x64-setup.exe`, choose More info and then Run anyway in the SmartScreen prompt, and start the app from the Start menu or a desktop shortcut. An MSI variant, `OpenWiki_X.Y.Z_x64_en-US.msi`, is also published, and the setup executable is the recommended one.

In both cases the last step of the first launch is the same: open Settings, go to the AI section and configure your provider. That is where the API key or OAuth login goes.

The unsigned Windows build is the detail to weigh if you work under a managed endpoint or a strict security policy, since a warning you are told to click through is exactly the pattern those policies look for.

## Building from source needs Tauri, Node 18 and a bundled markitdown converter

The development path is a Tauri project with a Vite and React front end. Prerequisites are Node.js 18 or newer, Rust on latest stable, macOS 13 or Windows 10 and 11, plus Xcode Command Line Tools on macOS and the Microsoft C++ Build Tools and WebView2 Runtime on Windows.

The clone and build sequence is:

```bash
# Clone the repo
git clone https://github.com/kdsz001/OpenWiki.git
cd OpenWiki

# Install dependencies
npm install

# Development mode
npm run tauri dev

# Build the app
npm run tauri build
```

Before producing release bundles, the bundled document converter has to be prepared, and that step differs by platform:

```bash
# macOS / Linux
./src-tauri/scripts/setup_markitdown.sh

# Windows PowerShell
./src-tauri/scripts/setup_markitdown.ps1
```

Skipping that step is the most likely way to get a broken release build, since the converter is bundled rather than fetched at runtime.

The npm scripts are minimal, which matches the setup: vite for development, tsc and vite build for a production bundle, eslint for linting, and a tauri script that wraps the Tauri CLI for both dev and build. The Tauri plugin set in the dependency list is what the desktop features ride on, including the clipboard manager, the updater, autostart, the shell and the process plugin.

## Two outbound readers ship enabled by default, and that is the trade

The privacy claim and the network behaviour of the app are both stated, and they need to be read together. URL reading can send the full URL to Jina Reader, and the Chinese UI can send foreign-language page text to Google Translate. Both are enabled by default for compatibility and can be disabled independently in Settings. Platform-specific readers may still contact the source platform or its API.

That is a fair disclosure, and the independent switches are the right design, but it does mean the default configuration is not fully offline. Anyone whose content cannot leave the machine has to change two settings before the first capture, not after a week of use.

Against the alternative of a plain clipboard manager or a hand-kept notes folder, the difference in approach is clear. A plain manager stores what you copied and stops there, with no provider key, no model in the loop and no outbound calls. OpenWiki gives you wiki pages, a graph, an ask sidebar and weekly reports in exchange for an API key, a model that organises your material on your behalf, and the two readers above.

For a reader who only wants the raw material, that trade is not worth it. For someone who has never kept a coherent notes system because the maintenance cost was the blocker, it is the reason to install the app.

On maintenance, the releases move quickly: v0.3.26 and v0.3.27 both landed on 2026-09-29 and v0.3.28 on 2026-10-07, matching the version string in package.json and the last push on 2026-10-07. The project is MIT licensed, and the knowledge base design was inspired by Karpathy's LLM Wiki idea.

## Conclusion

OpenWiki suits people who read and copy constantly, want their material turned into connected notes without a service subscription, and are comfortable giving an API key to a provider of their choice. It is the wrong tool if you need guaranteed offline behaviour, since URL reading and the Chinese interface can call Jina Reader and Google Translate, and the Windows build is unsigned. Verify first that the AI provider you pick is one of the three listed, that the two outbound readers are switched off in Settings if that matters to you, and that your content export path, one click to Markdown, gives you what you need before you let the model reorganise it.

## FAQ

### How does OpenWiki capture content on macOS and Windows?

Copying something opens a popup on the desktop that auto-dismisses after ten seconds, and only the content you actively choose to keep is saved. Text, images and URLs are supported, with source app detection, and you can trigger the popup manually with Cmd+Shift+C on macOS or Ctrl+Shift+C on Windows.

### Which AI providers does OpenWiki support and how do I sign in?

Anthropic (Claude), OpenAI and Google Gemini are supported, each with either an API key or an OAuth login, and you can pick a different model per provider. Gemini OAuth needs GEMINI_CLIENT_ID and GEMINI_CLIENT_SECRET in .env, and OpenAI OAuth needs OPENAI_OAUTH_CLIENT_ID.

### Does OpenWiki send my data to a server?

Your data is stored in a local SQLite database, but URL reading can send the full URL to Jina Reader and the Chinese UI can send foreign-language page text to Google Translate. Both are enabled by default and can be disabled independently in Settings.

### Can I get my captured notes out of OpenWiki?

Yes, one-click export to Markdown is part of the content management features, alongside filtering by type and time range and global search across both the content and the knowledge base.

### How do I install the OpenWiki desktop app?

Download the build for your platform from the releases page, the aarch64 or x64 .dmg on macOS, or the x64 setup executable or MSI on Windows. The macOS build is signed and notarized, while the Windows build is unsigned, so SmartScreen may warn on first launch.

## Sources

- [kdsz001/OpenWiki on GitHub](https://github.com/kdsz001/OpenWiki)
- [License: MIT](https://github.com/kdsz001/OpenWiki/blob/main/LICENSE)
- [Project website](https://openwiki.pages.dev)
- [README](https://github.com/kdsz001/OpenWiki/blob/main/README.md)
- [Releases](https://github.com/kdsz001/OpenWiki/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/kdsz001-openwiki