Examples

Every example below runs in the live demo, which loads a model once and lets you switch between them. Each page has the complete tools file — the same code the demo runs — plus the questions to try and an honest note on how small models handle it.

Example	What it shows
Basics	Three tools whose answers a model cannot guess — so a correct answer proves a real call happened.
JavaScript eval	A code interpreter in the tab: the model writes JavaScript, a Web Worker runs it.
This tab	Tools no server-side agent can have: they read and change the page you are looking at.
Crypto & encoding	SHA-256, base64 and UUIDs from Web Crypto — answers no model can guess, only compute.
Notes	Memory that survives a reload: three tools over localStorage.
Dates & units	Day counts, weekdays and unit conversions — the arithmetic small models get subtly wrong.
Live web	Live data from GitHub, Hacker News and Wikipedia — the model is local, only each tool’s request leaves the tab.
Just chat	No tools: an ordinary chat with the model running in the tab.
Decisions, Jev-style	Probabilities over your options from one forward pass — and a game of Snake run by a decision model.
Embeddings & RAG	Answer from your own documents, retrieved in the tab.

Every tools file follows one shape:

tool('name', 'What it does — the model reads this to decide when to call it.',
  { arg: 'string' },                 // shorthand JSON schema
  async ({ arg }) => ({ result }));  // any JSON back; the model answers from it

Keep each example to a handful of tools: a 0.5B model picks reliably from three names, not from thirty.