Rabbithole – an infinite, AI‑powered canvas for learning

What it is – Rabbithole is a lightweight web‑app that lets you create a branching “canvas” of documents that are answered by an LLM of your choice. You start with a document, ask a question at any point, and the answer spawns a new linked document, letting you explore ideas as a tree of self‑contained HTML pages.

Why it matters – It combines a zero‑setup front‑end with a flexible back‑end (the MCP server) that can talk to Claude Code, Codex, OpenAI‑compatible endpoints, or any custom model you run locally. All data stays on your machine – no accounts, no telemetry, no cloud storage.

---

Core concepts

Component	Role
Web app	Pure‑client side UI that renders the canvas and forwards prompts to a model endpoint you configure (OpenRouter, local OpenAI‑compatible servers, etc.).
MCP server	A small Node.js process that mediates between the web UI and “MCP clients” such as Claude Code or Codex. It runs locally, stores documents under ~/.rabbithole/ (or RABBITHOLE_DIR), and handles tool‑call routing.
Canvas	A series of HTML snapshots. Each snapshot contains the original content plus any follow‑up questions and the model’s answers, linked together so you can navigate forward and backward.

Getting started (quick‑start)

1. Web version – Open https://rabbithole.ing, paste a question, a URL, a Markdown file, or a PDF. The UI will send the prompt to the model you’ve configured.
2. Bridge to Claude Code / Codex – If you have the Claude Code or Codex CLI installed and signed in, run:

   npx @shlokkhemani/rabbithole bridge
   

   This prints a private pairing link and connects the browser page to the local CLI over the loopback interface.
3. MCP mode – Install the MCP client (Node 18+ required) and add Rabbithole as a tool:

   # Claude Code
   claude mcp add rabbithole -- npx -y github:shlokkhemani/rabbithole
   # Codex
   codex mcp add rabbithole -- npx -y github:shlokkhemani/rabbithole
   

   Then start an agent session and say “Open this document in Rabbithole.” The agent will keep the tool call pending while you interact with the canvas; the MCP server persists asks so you can disconnect and resume later.

Development & contribution

git clone https://github.com/shlokkhemani/rabbithole.git
cd rabbithole
npm install
npm run build   # builds the browser bundle
npm test        # runs the test suite

The repository ships pre‑built browser bundles, so downstream users installing the package from npm get a ready‑to‑run web app without needing to compile anything.

Documentation highlights

• Product & architecture tour – docs/01-tour.md
• Generated module & command maps – useful for developers extending the tool (docs/generated/*).
• Compatibility contract – defines how the web app talks to model endpoints and MCP clients (docs/compatibility.md).
• Design system – visual and interaction guidelines (docs/design-system.md).

License

MIT – see the LICENSE file.

---

Bottom line – Rabbithole is a self‑hosted, model‑agnostic, branching‑document UI that lets you turn any LLM interaction into a navigable knowledge map, with all data staying locally under your control.