Your models
Typillar generates code with a large language model, but it never runs that model on its own servers. Inference happens on your side, through your own Cloudflare Workers AI. Typillar is Cloudflare-native: the model is part of the same account that runs your app, its data, and its hosting — there’s no separate provider and no key to paste.
One model source: your Workers AI
Section titled “One model source: your Workers AI”Connect Cloudflare and the agent has its model. Workers AI runs on Cloudflare’s network under your account and token, and every token is billed to you. That’s the whole setup — see Models for choosing which model builds run on.
Why inference runs on your side
Section titled “Why inference runs on your side”- Cost transparency. You see and pay for inference directly, on your own account — no markup, no proxy, no surprise line item.
- One account, one bill. The model lives where everything else already does, so there’s nothing extra to connect or reconcile.
- A clean trust boundary. Because Typillar never runs the model, your prompts and generated code don’t pass through our inference servers. See Zero inference on our servers.
Real choice, still on Cloudflare
Section titled “Real choice, still on Cloudflare”Cloudflare-native doesn’t mean one model — there are three groups, all reached through your Cloudflare connection, and what separates them is what your account needs before you can reach them.
- Included, on any plan: a fast, cheap default for everyday builds, plus a smaller, faster model for lighter work like planning. These run against the free Workers AI allowance every Cloudflare account gets.
- Workers Paid, still on Workers AI, still billed as ordinary Workers AI usage — but Cloudflare gates two heavier open-weight coders behind the Workers Paid plan or prepaid credits. On a free account with no credits, Typillar refuses to reach for them rather than quietly building with something else.
- Closed frontier models — Claude, GPT, and Gemini — through Cloudflare’s AI Gateway unified billing: you pay each provider’s own rate on credits you load in your Cloudflare dashboard, with no Typillar markup.
The trust boundary is unchanged across all three: the request runs on your account, and nothing passes through our inference servers. See Models for the full catalog and how to choose one.
Every model in the catalog supports function calling, because the build loop is agentic and a model without tool calling can’t drive it. Attachment input is not universal: some models read images — each one that does is marked Sees images. on the Models page, and the ones that also read PDFs are marked Reads PDFs. — and the free defaults read neither.