@kanzo-tech/llm

The two doors to a model's answer, without React.

npm install @kanzo-tech/llm ai
npm install @browser-ai/web-llm @mlc-ai/web-llm   # for localModel, a model in the page

ai is a required peer, not an optional one: there is nothing here without it, and a host installs it once so that its own agent and @kanzo-tech/ai's components share one UIMessage and one transport. None of it is re-exported: import ToolLoopAgent, tool, Output and the message types from ai by name, as every package here does.

What it holds

import { generate, stream, LlmError } from "@kanzo-tech/llm";
import { localModel, localProvider, setLocalProvider, serveLocalModel } from "@kanzo-tech/llm/local";
import { scriptedModel } from "@kanzo-tech/llm/testing";

• stream({ model, system, prompt, output }) — an answer as it is written: streamText, never retried, read as text, elements (Output.array) or partial (Output.object), each of which throws what stopped the model once it ends. The SDK alone reports a failed stream to onError and ends it as if it had finished.
• generate({ …, providerOptions }) — the same, read whole: generateText, asked once. Every model call the library makes goes through one of the two.
• LlmError — llm/unavailable, data.reason "webgpu" (the page cannot run a model) or "unserved" (nothing serves it).
• localModel(id?), from @kanzo-tech/llm/local — the model, in the page, resolved by id through one provider registry: a host sets its provider once, setLocalProvider(localProvider({ baseURL: "/models", worker })), and every model listed in /models/app-config.json (WebLLM's AppConfig) is served; without an id, the first. WebLLM on WebGPU through @browser-ai/web-llm: both optional peers that only this subpath reaches. One engine per model per page; without WebGPU a call fails as LlmError, before anything is downloaded. A call may pass providerOptions.kanzo.grammar, an EBNF the engine holds every token to. loadLocalModel(id?) downloads and loads it before the first call, and localModelLoad(id?) / watchLocalModel(id, listener) say where that stands. serveLocalModel() is the worker's whole script.
• scriptedModel(script), from @kanzo-tech/llm/testing — the one model double: the AI SDK's MockLanguageModelV4 answering from a script of turns. A unit test passes it where a model goes; an end-to-end test registers the script with @kanzo-tech/testing's playwright(page, { models }), which registers it as that id's provider in the same registry.

Why it is not part of @kanzo-tech/ai

Reaching a model is not a user-interface concern, and not every caller draws one: a host's agent, a server route or a test needs the model and none of the components. @kanzo-tech/ai depends on this package and not the other way round.