Build with llamay
Live playground
This page talks to a llamay running on your own machine. Nothing is sent to
llamay.com — the requests go from your browser to
127.0.0.1:11435 and the tokens are generated by your hardware.
If nothing is running, the page says so rather than pretending.
Checking for a local llamay…
Two things have to be true.
A server has to be running:
llamay serve -cors https://llamay.com
And it has to be told this page may call it. That is what
-cors does, and it is off by default on purpose: a server on
loopback is reachable by every page you have open, and the browser's refusal
to make the call is the only thing between it and any site that would like a
free model. The flag names one origin and grants nothing to any other.
Needs llamay 0.1.55 or later. On an older build the flag does not exist and the page cannot reach the server, however it is started.
…The same request, as curl
What each recipe shows
Streaming
Tokens arrive as server-sent events and render as they land. The first frame carries the role and no text, which is what lets a client paint an empty bubble immediately.
Structured output
A JSON Schema is enforced while sampling, not checked afterwards. The
reply parses or the request fails — and an enum cannot be
stepped outside, because the sampler never sees those tokens.
Stop sequences
Generation halts at the delimiter and the delimiter is not in the answer. A sequence split across two tokens still stops, which is the case that makes this harder than it looks.
Contexts
A prefix is computed once and forked per request, so the second call pays for its own tokens only. Watch the reported prefill drop on the second run.
Where the text goes
Into your browser's memory and out to 127.0.0.1. This page has no
backend of its own: there is no endpoint on llamay.com that sees what you
type here, because every request is issued by the tab to your own machine.
Turn the network off after loading the page and everything below still works.
Where to go next
Recipes has each of these as a complete script, plus sixteen more. The HTTP API documents every field the playground sets. Deployment patterns covers what happens when it is not your laptop making the call.