Learn something while your AI thinks.
A terminal companion for the 5–10 minutes you spend waiting on Claude Code, Cursor, or any long-running agent. Type a topic, pick a level, and drip shows you one-sentence insight cards — one at a time, one on screen. A new drip replaces the last every 10 seconds. Glance over when your brain has a second.
One file. Zero runtime deps. Node 18+. MIT.
Install and run directly from GitHub — no npm registry needed:
npx github:ianfh0/drip # interactive onboarding
npx github:ianfh0/drip "diffusion models" # topic given, picks level
npx github:ianfh0/drip "rust borrow checker" -l advanced # fully specified
npx github:ianfh0/drip "Thinking, Fast and Slow" # a book title works as a topicOr clone it:
git clone https://github.com/ianfh0/drip.git
cd drip
./drip.jsThe first time you run drip, it walks you through Claude auth in 30 seconds (see below). After that, it just starts.
Drips auto-play every 10 seconds. You don't have to press anything.
| Key | Action |
|---|---|
| space | skip ahead — don't wait for the 10s tick |
| p | pause / resume auto-play |
| q | back to the topic picker |
| ctrl+c | exit drip |
One card at a time, one good rhythm. No feed, no scroll, no settings.
First time you run drip, it asks one question and walks you through setup. No commands to memorize. Two paths:
drip installs @anthropic-ai/claude-code for you, probes whether you're already signed in (common — just continues), and otherwise launches Claude Code once for you to sign in via browser. Then drip starts automatically.
drip opens the API key page in your browser. You paste the key at the prompt, drip offers to save it to your shell rc (~/.zshrc / ~/.bashrc / fish config) so it persists, and drip starts automatically.
drip never sees or stores your key. Everything is client-side. No backend, no proxy, no accounts, no telemetry. The package has zero runtime dependencies — if this repo disappeared tomorrow, your existing install would keep running.
Cost reality for the API key path: Sonnet with prompt caching lands around $0.003 per drip. A full hour of 10-second auto-play (360 drips) is about $1.
One sentence. 8–10 words. No title, no bullets. Concrete — a number, a name, a mechanism, or a surprise. Stands alone.
GPT is matrix multiplication plus a nonlinearity, stacked 50 times.
Models generalize thousands of steps after overfitting — nobody knows why.
Wild sloths mostly die descending trees to poop each week.
System 2 only engages when System 1 flags something weird.
Reading time for a fast reader: about 2 seconds. Tuned to be absorbed at a glance, not to demand your attention.
- beginner — plain English, zero jargon. Vivid mental models, not definitions. Written for a curious friend who's heard of the topic but never studied it.
- intermediate — assume basics are known. Named concepts, specific numbers, mechanisms, trade-offs. Written for a practitioner warming up.
- advanced — senior-practitioner depth. Named effects, specific researchers, recent developments, surprising edge cases. Written for someone who wants the interesting stuff.
Depth also progresses within a session — drip #1 is foundational, drip #50 reaches the frontier. Each drip stands alone, but they flow like a smart friend's walk-through of the topic.
- Instant advance. A lookahead buffer pre-generates the next drip in the background the moment the current one lands. When the 10-second tick fires (or you press space), the next drip is already sitting there — no loading, no spinner.
- Streaming. The very first drip of a session streams in live, typewriter-style. Everything after that is pre-generated and displays instantly.
- Prompt caching. The system prompt is identical across every drip, so Anthropic caches it — 90% off on the cached portion after the first call.
- Idle pause. No input for 90 seconds and drip stops generating. Resumes on the next keypress. Your API budget doesn't burn while you're tabbed out.
- No repeats. A rolling list of already-shown drips goes back to the model, so it doesn't restate itself.
Sonnet by default — best quality-per-token for short insight cards. Override:
DRIP_MODEL=haiku drip "GPU architecture" # cheaper, faster, slightly shallower
DRIP_MODEL=opus drip "category theory" # slower, denserA weekend hack for anyone waiting on AI. MIT, zero-dep Node. No infra, no backend, no keys drip sees, no accounts, no telemetry, no waitlist. Fork it, extend it, ignore it.
Maintained: best effort. No roadmap. PRs welcome; no promises.
drip/
├── drip.js the whole CLI
├── package.json bin entry, Node 18+
├── README.md this file
└── LICENSE MIT