Free · open source · local-first

chat is a straight line.
your curiosity isn’t.

Rabbithole turns documents, questions, and answers into a branching canvas. Select what catches you, ask, and follow the answer as far as it goes.

Open the browser app Install for your agent Star on GitHub

select anything, ask anything.

Two ways in

use the AI you choose

Open the hosted app immediately, or connect Rabbithole to the terminal agent already doing your work.

01no install

Use it in your browser

Already pay for Claude or ChatGPT? Run npx @shlokkhemani/rabbithole bridge and click the link it prints — answers come from Claude Code or Codex on your machine, on your plan. Or bring an OpenRouter key, or point Rabbithole at an OpenAI-compatible local model such as Ollama. Your holes and settings stay in this browser.

Bridge requests stay on your machine. OpenRouter requests go directly to OpenRouter. Local-model requests go only to the endpoint you configure.

Open Rabbithole
02Claude · Codex · MCP

Connect your AI agent

Install the MCP server and let your existing agent answer. Rabbithole supplies the local canvas, branching documents, and durable storage.

The server and canvas run on your machine. Documents and questions go only to the agent you already use.

See the one-line install

How it works

every question leaves a trail

01

Open

Start with a PDF, Markdown file, article, or a question of your own.

02

Ask

Select the exact words that snag you. Ask freely or use a learning lens.

03

Descend

The answer becomes a child document. Select again and keep branching.

Agent install

connect in about a minute

Requires Node 18+ and a browser. No clone or build is needed for the MCP server.

Claude Codeone line
claude mcp add rabbithole -- npx -y github:shlokkhemani/rabbithole
Codexone line + timeout
codex mcp add rabbithole -- npx -y github:shlokkhemani/rabbithole
Full MCP instructionsConfiguration, timeouts, other MCP clients ↗ Run the browser app locallyClone, build, self-host, and optional proxy ↗