Plenty of tools will draw a mind map. What this does differently is type the nodes and then make one type impossible to miss: blockers come out deep red, so a long conversation about a product direction resolves into where you are actually stuck. Because Claude does the extraction itself there is no second model bill and no third-party service holding the conversation.
A Node server plus a local web canvas. Claude analyses the conversation itself and writes structured nodes into a map you open in the browser; there is no second API key because the analysis is done by the model you are already talking to.
- A conversation turned into structured nodes on a map — `capturemind`
- Node types that carry meaning, with `blocker` rendered in deep red so bottlenecks are visible at a glance; the rest are `root`, `goal`, `problem`, `idea`, `insight`, `action`, `question` and `note`
- Topics created, listed, fetched and deleted — `create_topic`, `list_topics`, `get_topic`, `delete_topic`
- Nodes and edges added, edited, deleted and imported in bulk — `add_node`, `update_node`, `delete_node`, `add_edge`, `delete_edge`, `batch_add`
- The map re-arranged into a hierarchical tree, searched full-text, and exported as Mermaid, JSON or Markdown — `layout_topic`, `search`, `export_topic`
Node.js 18+ and Claude Desktop. Clone the repository, `npm install`, `npm start` — the canvas is at `http://localhost:3001` — then point Claude Desktop at `mcp-server.js` with the full path. No extra API key is needed. The interface switches between Chinese and English and remembers the choice.