A WYSIWYG markdown editor just went full Web4—agents can write in your notes while you watch the CRDT sync in real-time.
The Summary
- OpenKnowledge launches as an open-source, AI-native markdown editor with direct integrations for Claude, Codex, and Cursor to edit files side-by-side with users
- Built on a dual-observer CRDT that keeps ProseMirror's visual editor and raw markdown perfectly synced, with git-based collaboration that shows what agents are editing in real-time
- Ships as MacOS app, web UI, and CLI with built-in MCP servers, RAG capabilities, and embedded terminal—designed for "AI Second Brain" workflows where agents write alongside humans
The Signal
OpenKnowledge is the first serious editor built for the assumption that agents will be co-authors, not assistants. The architecture makes this explicit: integrations let Claude, Codex, and Cursor open an OpenKnowledge editor inside their own apps. You're writing in one window, the agent is writing in another, and both of you see changes propagate through a CRDT in real-time. This isn't chat-to-generate-text. This is shared workspace where both parties have write access to the same document state.
The engineering reveals how much infrastructure it takes to make this feel simple. ProseMirror—the editor framework—uses abstract syntax trees that don't preserve exact byte positions. Markdown is plain text where every character matters. The team built a bidirectional converter that goes from ProseMirror AST to markdown without information loss. Then they wrapped both representations in a dual-observer CRDT so edits in either format update the other instantly. This is the plumbing Web4 needs: tools that let agents manipulate structured data the same way humans do, with conflict resolution baked in.
"The CRDT + git also power a collaborative experience that shows what Agents are doing in the markdown, have undo/redo, and version history."
The "share" feature is GitHub under the hood. Every document is a git repo. Collaboration is pull requests without the ceremony. Agents commit changes to the same history humans do. This makes OpenKnowledge one of the few tools where agent edits aren't ephemeral suggestions—they're durable contributions with full audit trails. You can see what the agent wrote, when, and roll it back if it went sideways.
The tech stack matters here:
- Tiptap/ProseMirror for the visual editor
- Yjs for CRDT syncing across multiple editors
- Orama for local search and RAG
- Electron for the MacOS app shell
- Model Context Protocol (MCP) servers built-in for agent integration
That last piece is key. MCP is Anthropic's spec for how agents discover and use tools. OpenKnowledge ships with MCP servers, which means agents can query your notes, update files, and pull context without custom integrations. This is infrastructure for the pattern where agents maintain living documentation while you work on something else.
The Implication
If you're building agent workflows, watch this project. The dual-CRDT architecture solves a real problem: how do you let agents edit structured content without flattening it to text prompts and responses. OpenKnowledge's answer is to give agents the same primitives humans use—a real editor, version control, conflict resolution—and then get out of the way.
The open-source release means the patterns here will spread. Expect more tools where agent edits show up as commits, where CRDTs keep human and machine state synced, where git becomes the collaboration layer instead of a bolt-on. This is what native Web4 tooling looks like: built from the start for two kinds of writers.