Rabbithole

An infinite canvas for learning. Open a document, select any text, ask a question — and the answer opens as a fully-rendered child document. Follow whatever pulls at you, as deep as it goes. Every hole is saved and revisitable.

🌐 rabbithole.ing

There are two ways in:

• The web app — rabbithole.ing. Bring an OpenRouter key or point it at a local model. Static site, no account, no backend: your key stays in your browser, and so do your documents.

• The MCP server — for terminal agents. Claude Code, Codex, or any MCP client does the answering; Rabbithole gives it a canvas in your browser. The server, storage, and canvas all run on your machine — your documents and questions go only to the agent you already use.

Use it on the web

Open rabbithole.ing and start from anywhere: drop in a PDF or Markdown file, paste a URL, import a .rabbithole or snapshot .html — or just ask a question and let the answer become your first document.

Two ways to run a model:

• OpenRouter (recommended) — one key, every major model. The model picker pulls OpenRouter's live catalog.

• Local — any OpenAI-compatible endpoint: Ollama, LM Studio, llama.cpp. No key required.

Keys never leave the browser: they're stored locally (or session-only, your choice) and sent exclusively to the provider origin you configure. Exports scrub anything credential-shaped.

Run the browser version locally

Requires Node 18+:

git clone https://github.com/shlokkhemani/rabbithole.git
cd rabbithole
npm install
npm run build
npx -y serve web/dist

Open http://localhost:3000 (or the URL printed by serve). The local browser build has the same OpenRouter and OpenAI-compatible local-model options as rabbithole.ing, and its documents and provider settings stay in that browser's local storage.

Holes persist in IndexedDB, and each document gets a memorable local URL such as rabbithole.ing/curious-teacup-abcdef. That path names a record in your browser's database — it is not a sharing link. To move a hole between machines, export the .rabbithole file; to share something readable, download a snapshot.

Self-hosting is static: run npm run build and serve web/dist from any host. The optional workers/fetch-proxy Cloudflare Worker enables URL ingestion for sources that block browser CORS (set RABBITHOLE_PROXY_URL at build time to point the app at your relay). Serve index.html as the fallback for unknown single-segment paths so document URLs survive a direct visit or refresh.

Related MCP server: team-docs-mcp

Quick start

Requires Node 18+ and a browser. Pick your agent:

Claude Code

claude mcp add rabbithole -- npx -y github:shlokkhemani/rabbithole

Codex

codex mcp add rabbithole -- npx -y github:shlokkhemani/rabbithole

Then raise the tool timeout in ~/.codex/config.toml — Codex's 60-second default is shorter than Rabbithole's blocking wait, and codex mcp add cannot set this field:

[mcp_servers.rabbithole]
command = "npx"
args = ["-y", "github:shlokkhemani/rabbithole"]
tool_timeout_sec = 600

Any other MCP client — most accept this shape in their MCP config:

{
  "mcpServers": {
    "rabbithole": {
      "command": "npx",
      "args": ["-y", "github:shlokkhemani/rabbithole"]
    }
  }
}

Then start a new agent session and say:

open this document in rabbithole

A browser tab opens. Select any text on the page, ask a question (or tap a lens — Explain · ELI5 · Example · Go Deeper), and the answer streams in as a new document. Keep going.

If a host ever reports a tool timeout, nothing is lost — questions are saved and re-queued the next time the agent listens.

git clone https://github.com/shlokkhemani/rabbithole.git
cd rabbithole && npm install
claude mcp add rabbithole -- node "$(pwd)/bin/mcp-server.js"

🤖 If you are an AI agent setting this up

Follow these steps exactly:

1. Check Node: node --version must be ≥ 18.

2. Register the server with the one-liner above that matches the agent you are running inside (claude mcp add … for Claude Code, codex mcp add … for Codex, or write the JSON block into the client's MCP config file). The npx -y github:shlokkhemani/rabbithole form needs no clone and no install step — npx fetches this repo and runs bin/mcp-server.js over stdio. For Codex, also add tool_timeout_sec = 600 under [mcp_servers.rabbithole] in ~/.codex/config.toml; codex mcp add cannot set that field.

3. Verify: after the client reconnects, three tools should be available — open_rabbithole, answer_branch, list_rabbitholes. (In Claude Code, claude mcp list should show rabbithole as connected. The server itself also responds to a standard MCP initialize with server name rabbithole.)

4. Tell your human what to do next: they should start a session and say "open <some document> in rabbithole" — you then call open_rabbithole with { title, content } (or { title, file_path }; add base_url when the markdown came from a URL or repo; add assets for local images referenced as asset:name.png), which blocks until they select text and ask. Answer with answer_branch (stream chunks with partial: true, finish with a titled final call) and keep looping until status='session_closed'. If a long wait returns status='keep_listening', immediately call open_rabbithole with the returned { hole_id } and do not re-send content.

Notes for you: the tool call blocks by design (long-poll) — that is normal, not a hang. The first npx run fetches this repo, so allow it time; later runs are cached. If the browser must not auto-open (headless), set RABBITHOLE_NO_BROWSER=1 in the server's env.

Tools

The loop: open_rabbithole → branch_request → answer_branch → branch_request → … → session_closed. Long waits may return keep_listening; immediately call open_rabbithole again with the returned hole_id. If the host reports a tool timeout, do the same — questions are saved.

For research PDFs, page renders are the dependable figure source. For arXiv links, prefer fetching the HTML version and opening that content with base_url instead of ingesting the PDF.

What's inside

• Reader mode (default): fullscreen reading, branches sidebar, breadcrumbs; selections become inline marks (pending → ready); clicking a mark jumps to its answer; child docs carry a FROM strip that jumps back to the exact origin.

• Streamed answers: words appear live with a breathing caret — in the reader, the thread, and the canvas card.

• Rich Markdown: answers can use math, highlighted language code fences, show diagrams, URL-based resolution for relative links/images, and local image assets via asset:name.png; source stays as Markdown for copy/export, while frozen snapshots inline assets into the HTML.

• Interactive checks: answer multiple-choice questions inline; progress survives reloads and portable backups, while shared snapshots start clean.

• Lenses: one-tap presets on the ask popup — Explain · ELI5 · Example · Go Deeper (keys 1–4).

• Follow-up chat: a composer under each document asks about the doc as a whole; answers render inline and are branchable like any other text.

• Canvas mode: infinite pan/zoom, draggable/resizable cards, edges that attach to the exact selected text in the parent, collapse, auto-layout.

• Navigation: j/k walk marks, ↵ opens, ⌫ jumps back up, ⌘K searches the whole hole.

• Share/export: copy any trail or document as Markdown; Download snapshot produces a single self-contained .html — data, assets, and a read-only client in one file anyone can open; Export Rabbithole (web app) produces a .rabbithole backup for device transfer — MCP holes are already plain JSON on disk; or ask the agent for a synthesis of the whole journey.

• Durable asks: questions asked while the agent is away are saved and re-queued on resume — the agent answers them first thing.

• Persistence: holes auto-save as JSON under ~/.rabbithole/; resuming restores the doc, scroll position, mode, and canvas framing.

The MCP host stores each hole as a JSON file directly under ~/.rabbithole/ (RABBITHOLE_DIR overrides the base directory) and assets under the matching asset directory. The web .rabbithole file is the same persisted hole JSON wrapped as { format: "rabbithole", format_version: 1, hole, assets }, with assets base64-encoded into the single JSON file for portability.

Configuration

Repo layout

• bin/mcp-server.js — entry point (stdio MCP server)

• src/core/ — host-independent document engine, rendering, artifacts, and contracts

• src/ui/ — shared live/frozen browser runtime

• src/node/ — MCP host, filesystem storage, local HTTP/SSE, and PDF ingestion

• src/web/ — static BYOK browser host and IndexedDB storage

• build.mjs — builds the committed MCP bundles and the static web app

• dist/ — committed browser bundles used by GitHub npx installs

• web/dist/ — generated static web app (untracked build output)

• scripts/ — reproducibility checks and publish assembly

• test/ — capability-oriented unit, contract, integration, end-to-end, performance, and packaging suites

• workers/fetch-proxy/ — optional allowlisted URL-ingestion relay

• website/public/ — public deployment assets consumed by build:publish

Development

See CONTRIBUTING.md for the development workflow and ARCHITECTURE.md for system boundaries. Compatibility, testing, and interface rules live under docs/. The browser runtime source lives in src/ui/ and is bundled into committed artifacts under dist/. When editing the UI, run:

npm run build
npm run check:dist

Commit both the source changes and dist/. There is no prepare build step; GitHub npx installs use the committed artifacts.

Production deployment

The Deploy Cloudflare Pages workflow runs the complete test suite and deploys publish/ to the rabbithole Pages project on every push to main. It can also be rerun manually from GitHub Actions. Each Cloudflare deployment is tagged with the exact Git commit.

The workflow requires:

• repository variable CLOUDFLARE_ACCOUNT_ID;

• repository secret CLOUDFLARE_API_TOKEN, scoped to Account → Cloudflare Pages → Edit.

Configure them with gh variable set CLOUDFLARE_ACCOUNT_ID and gh secret set CLOUDFLARE_API_TOKEN; both commands prompt without committing credentials.

License

MIT