TablaCognita

Co-write with any AI. In your browser. No data on anyone's server.

TablaCognita is a browser-based Markdown editor where AI is a first-class participant, not a feature bolted onto a text box. Select text and send it to your AI. Watch edits land in real time. Undo anything. The AI sees the document through MCP tools — the same protocol that powers Claude Desktop, Cursor, and every other MCP-capable client.

Your document lives in your browser. The relay server is a stateless wire. Nobody stores your writing.

How It Works

┌──────────────┐   MCP (stdio | HTTP)   ┌──────────────┐   WebSocket   ┌──────────────────┐
│  Your AI     │◄───────────────────────►│    Relay      │◄────────────►│  Browser Editor   │
│  (Claude,    │   Tool calls/results    │  (stateless)  │              │  (source of truth)│
│   Cursor,    │                         │               │              │                   │
│   any MCP)   │                         └──────────────┘              │  CodeMirror 6     │
└──────────────┘                                                       │  Live preview     │
                                                                       │  Section tracking │
                                                                       │  Lock management  │
                                                                       └──────────────────┘

The AI reads and writes through 17 MCP tools. The editor handles section addressing, fuzzy text matching, collision avoidance, and undo. The relay just passes messages. That's it.

Quick Start

git clone https://github.com/pstryder/tablacognita.git
cd tablacognita
npm install
npm start

Open http://localhost:3000 in your browser. The editor is live.

Connect your MCP client. For Claude Desktop, add to your config:

{
  "mcpServers": {
    "tabla-cognita": {
      "command": "node",
      "args": ["server/index.js", "--transport", "stdio"],
      "cwd": "/path/to/tablacognita"
    }
  }
}

Your AI can now read your document, edit sections, replace text, and respond to selections you send it from the editor.

What Your AI Can Do

Read — read_document, get_structure, get_section. The AI sees your full document, section outline, or individual sections by heading name or stable ID.

Write — replace_section, replace_text, insert_after, append. Section-level edits with auto-locking. Fuzzy text matching that works through markdown formatting (**bold text** matched by searching "bold text"). Full document replacement requires your confirmation in the browser.

Coordinate — request_edit_lock, release_lock, get_cursor_context, get_dirty_regions, poll_context. The AI knows where your cursor is, which sections you've edited since it last looked, and can lock sections for multi-step operations. All locks have TTLs — nothing stays locked forever.

Manage — open_document, snapshot, restore_snapshot, get_revision_history. Named checkpoints saved to your browser's IndexedDB. Restore any snapshot with one tool call.

Key Design Decisions

Browser is source of truth. The relay holds zero document state. If the relay restarts, your document is still in your browser. The AI reconnects and continues.

Section IDs are stable. Rename a heading from "Introduction" to "Overview" — the section ID doesn't change. The AI can always find its way back.

Agent edits are undoable. Every AI edit goes through CodeMirror's transaction system. Ctrl+Z works on AI edits exactly like your own.

Locks are intent signals, not mutexes. Auto-locks during edits are atomic and invisible. Explicit locks for multi-step operations have TTLs and never block you — just a gentle indicator that the AI is working on a section.

Notifications are fire-and-forget. The editor tells the AI when you move between sections, rename headings, or delete content. If the transport doesn't support notifications, the AI

... View full README on GitHub