OpenKnowledge Turns a Markdown Folder into a Smart Knowledge Base for AI Agents

I've been keeping work notes and architectural specs in plain Markdown files inside repositories for a while now. On one hand, it's convenient: everything is stored locally, versioned with git, and opens in any IDE. On the other hand, writing lengthy texts in raw Markdown with raw tags and tables quickly becomes tedious. You want a proper visual editor on the level of Notion or Google Docs, but without being tied to someone else's cloud and proprietary data formats.

The situation got more complicated when Claude Code, Cursor, and other terminal agents became part of my daily workflow. You constantly have to feed them context, ask them to rewrite a section, or update documentation based on the codebase.

Recently, the Inkeep team released OpenKnowledge as open source. Essentially, the authors tried to combine a visual document editor with a working environment for neural network agents.

Under the Hood

OpenKnowledge opens any local folder on disk containing .md or .mdx files. This could be your repository root, a folder with technical documentation, or even an old Obsidian vault. There are no proprietary databases here: the app reads files from disk and saves changes back in plain format.

At the same time, you get a full-featured WYSIWYG. Headings, lists, links, interactive tables, and embedded HTML blocks are formatted on the fly as you type. Your eyes don't stumble over markup syntax, and the resulting file remains standard Markdown.

The main feature is hidden on the side. The editor has a built-in panel for interacting with language models via Model Context Protocol (MCP) and local CLI tools. The app can work side by side with Claude, Codex, OpenCode, Pi, and other agents.

Main Editor Features

If you break down the project into components, you get an interesting set of features:

1. Honest WYSIWYG without compromises. You edit the document in a convenient visual view, and the app saves valid Markdown or MDX on the fly. Visual components, a graph of links between pages (wiki links), tabs, and quick file search are all supported.

2. Collaborative editing with AI. The agent opens in the second half of the screen and sees the structure of your documents. You can ask the model to create an architectural plan, find inconsistencies in a spec, or rewrite a section while watching the changes in real time.

3. Ready-made integration protocols. On first launch, the utility automatically detects agents installed on your system (Claude Desktop, Cursor, Claude Code) and sets up MCP servers and search skills for your local knowledge base.

4. Synchronization and team collaboration via Git. The project doesn't impose paid team subscriptions. Synchronization between devices and colleagues relies on plain Git and GitHub.

How to Install and Run

The project has two usage options: a ready-made desktop app or a local web interface that runs via Node.js.

Desktop builds for macOS (Apple Silicon), Windows 10+, and Linux (deb/rpm) are available on the project's official website and in the repository releases.

If you're working on a server, on an Intel Mac, or simply prefer launching tools via the console, a CLI client is available. It requires Node.js version 24 or higher:

# Устанавливаем CLI глобально
npm install -g @inkeep/open-knowledge

# Переходим в папку с документами или проектом
cd your-project

# Инициализируем проект и подключаем найденные AI-агенты
ok init

# Запускаем локальный веб-сервер и открываем браузер
ok start --open

The ok init command walks through the system, checks for installed assistants, and sets up the connection between them and the editor.

Where This Comes in Handy in Practice

The authors position the tool for several tasks right away.

The first obvious scenario is technical documentation and architectural specs. When you need to describe a system design, write an RFC, or prepare a report, the visual interface saves time. At the same time, the agent can read neighboring code files via MCP in parallel and suggest accurate function signatures.

The second scenario is a personal knowledge base or "second brain." If you've been keeping notes on the LLM Wiki concept (which Andrej Karpathy often talks about), where an agent helps organize and link disparate thoughts, OpenKnowledge takes on the role of the interface to this base. Full-text and semantic search help quickly retrieve needed fragments through prompts.

The Bottom Line

OpenKnowledge looks like a successful attempt to lift Markdown editing out of the spartan conditions of plain text editors without locking users into closed cloud services.

The code is distributed under the GPLv3 license, written in TypeScript, and is actively growing its community. The downside is the requirement for Node.js 24+ for the web version, which may require updating the environment on older machines. But if you frequently work with local specs and want to give your AI assistants convenient access to documents, it's definitely worth trying the project.