Start with the tool in front of you
On Tinker & Tokens, Terminal Twin lets you inspect saved model responses. Patch searches site material and returns prepared guidance. Those tools do not generate a new model response or call a paid inference service when you use them. A saved example can still teach a useful habit: compare the answer with the information the prompt actually supplied.
A live AI tool works differently. It generates a new response using computing power somewhere. Before treating a tool as free, ask where that work happens, what has to be downloaded, and what happens when its allowance runs out.
A model can work on your own device
Some browser features can run a model on the visitor’s computer. Chrome’s Prompt API is one example. The browser manages the model, and the site can ask it to generate text locally. This is an option for a compatible desktop setup, not a promise that every phone, tablet or computer can run it.
The browser may need a substantial model download and enough free storage, memory and processing capacity. Running it uses battery or electricity, and an internet plan may charge for downloads. Browser updates, model downloads and telemetry are separate from where a particular prompt is processed.
Tinker & Tokens is researching this approach. The working tools described above remain saved examples and prepared site search; this guide does not activate or download a model. Any future experiment should ask clearly before downloading and keep a usable fallback.
A hosted free plan is an allowance
Hosted inference runs on someone else’s servers. A provider may include a limited free allowance and refuse further requests when it is exhausted. For example, Cloudflare’s Workers AI documentation describes a daily Free allocation and errors after the limit. Paid plans have different billing rules. Some models require paid billing even when other models have a free allowance, so check the exact model as well as the account plan.
An API is the connection a website uses to request that work. Its bill normally follows the service account behind the connection. “Free to the visitor” therefore does not tell you what the website owner pays. Current plan, quota, billing enrollment and every fallback service matter; do not assume a free chat-app subscription pays for an API.
No inference bill does not mean no costs
Even a static site must deliver its pages, pictures and videos. If it also serves a large model file, every uncached download adds traffic. As an illustration, a 250 MB file downloaded one million times totals 250 TB in decimal units. That is arithmetic, not the measured size or audience of a model on this site.
Hosting described as unmetered can still have resource limits and acceptable-use rules. Spaceship publishes such limits for its shared hosting. A visitor’s local model does not turn the website into a service with unlimited bandwidth or capacity.
For tinkerers: draw a real spending boundary
Keep service credentials out of browser code. Anything sent to a visitor’s browser can be inspected by that visitor. If you later build a hosted tool, put credentials and enforcement on a server you control, verify that the provider truly stops at its free limit, and avoid a silent paid fallback.
A disabled button or a JavaScript request counter is not a spending control: another program can call the endpoint directly. A bot challenge may help reduce abuse, but its result must be checked by the server. Cloudflare explicitly requires that validation for Turnstile. Even a valid challenge does not impose a billing ceiling.
Protect an expensive generation action without blocking ordinary articles, search crawlers and verification files. Treat model output as untrusted text rather than executable code or instructions to visit a link. A local model can make mistakes just as a hosted model can.
Try a useful check without buying anything
Open a saved Terminal Twin example and identify one fact the answer could not know from the prompt. That exercise needs no new model account or download. If you already use an AI app, try fictional notes instead of private school, workplace or customer information.
Before enabling any new live tool, look for four clear answers: where it runs, whether it downloads anything, what its limits are, and how to stop it. Keep a copy of important source material and check the response before acting on it. “Local” and “free” describe parts of the setup; neither guarantees accuracy or complete privacy.