prisma/mrkdwn

mrkdwn is a collaborative writing workspace, think Notion or Google Docs, where humans and AI agents work together in the same document at the same time. You get live markdown editing with rich features like tables, task lists, and emoji, plus real-time cursors and comments. The standout feature is a one-click "invite your agent" button that brings an AI like Claude Code into the doc as a co-author with its own avatar, cursor, and ability to leave comments. When you invite an agent, it joins through a simple API that uses a familiar "old text, new text" editing pattern, the same approach AI coding tools already use when editing files. The server then applies those changes as small, staggered chunks so you actually watch the agent type alongside you, rather than having text suddenly appear. Meanwhile, every edit from both humans and agents syncs through a conflict-resolution system (called Automerge) that ensures nobody's work ever gets overwritten, even if two people edit the same sentence simultaneously. Everything is continuously backed up to cloud storage, with a status indicator showing green when saved, amber when pending, and red if the connection drops. This is built for teams or individuals who already use AI coding assistants and want them participating in documents, not just code. A product manager could draft a spec, mention an agent to fill in technical details, and watch it contribute in real time. A developer could leave unchecked to-dos in a doc and an agent would pick them up when it joins. The project also includes canvas boards, visual spaces with sticky notes and connecting lines, that agents can read and modify alongside humans. What's notable is how compact this is: roughly 3,000 lines of TypeScript running as a single process. There's no Kubernetes, no serverless complexity, no separate websocket infrastructure. It's essentially a demonstration of Prisma's platform for building real-time, stateful applications that combine web interfaces, live collaboration, and AI agents, all in one deployable unit that starts with just two commands.