Inkeep, a Y Combinator-backed documentation AI startup, released OpenKnowledge on June 27: a free, GPL-licensed WYSIWYG markdown editor that ships built-in MCP servers for Claude Code, OpenAI Codex, and Cursor. The tool lets AI coding agents read, search, and rewrite a user’s local markdown files without routing any data through a cloud server.
The product landed on Hacker News as a “Show HN” post on launch day, according to TechTimes.
How the Agent Integration Works
When a user runs ok init from the command line, OpenKnowledge detects which AI agent harnesses are installed locally and writes configuration files for each, including a pre-built MCP server that exposes vector search, wiki-link navigation, and document-editing capabilities. A Claude Code session can then access and edit the entire knowledge base as a first-class file system, with no manual setup required.
The MCP server includes what the team calls “agentic search,” combining full-text and vector-based retrieval powered by Orama, an open-source TypeScript search engine that runs inside any JavaScript runtime without external dependencies. Agents can answer natural-language queries against the knowledge base, retrieving notes related to a concept even when the user used different terminology across documents.
The CRDT Layer
What distinguishes OpenKnowledge from prior attempts to put AI inside a note-taking application is the editing layer. The tool uses a dual-observer CRDT architecture built on yjs that holds a ProseMirror rich-text document and its raw markdown representation in continuous, lossless synchronization. Co-founder Nick Gomez described the approach in the Hacker News thread as two independent observers monitoring the ProseMirror state and the markdown state separately, applying delta transformations to keep them in sync whenever either changes.
The practical effect: Claude Code or Codex can write directly to a markdown file and see those changes reflected immediately in the WYSIWYG editor, and vice versa, with byte fidelity preserved in both directions. No server is required to coordinate merges.
The Local-First Constraint
The product enforces a strict local-first architecture. All data stays on the user’s machine. The team collaboration layer uses CRDTs rather than a central server to resolve conflicts, meaning two users, or one user and an AI agent, can edit the same document simultaneously without locking or data transmission to third parties.
For teams operating in regulated environments or handling sensitive documentation, this is the architectural distinction that matters: the AI has access to the knowledge base, but the knowledge base never leaves the local filesystem.