@puckguo123/llm

LLM adapters for OpenAI-compatible endpoints, Anthropic, and a scripted mock. Zero dependencies.

The single StreamFn seam lets you swap providers without touching your agent code.

Install

npm install @puckguo123/llm @puckguo123/core

Subpaths

| Subpath | Use when | |---|---| | @puckguo123/llm | High-level getModel() / createStreamFn() registry (uses your login keys) | | @puckguo123/llm/openai | Direct OpenAI-compatible client (e.g. DeepSeek, Moonshot, Ollama) | | @puckguo123/llm/anthropic | Anthropic native client | | @puckguo123/llm/mock | Deterministic scripted responses for tests |

Quick start

import { Agent } from "@puckguo123/core";
import { createStreamFn, getModel } from "@puckguo123/llm";

const model = getModel("deepseek-chat");
const agent = new Agent({
  stream: createStreamFn(model),
  systemPrompt: "You are a careful assistant.",
});

const { text } = await agent.run("Summarize the news today");

Mock for tests

import { createMockStreamFn } from "@puckguo123/llm/mock";

const stream = createMockStreamFn([{ text: "hello" }, { text: " world" }]);

The mock yields a deterministic script — perfect for CI without any network.

Provider discovery

@puckguo123/llm re-exports the credential store and live GET /models discovery from @puckguo123/llm/auth. After puck /login anthropic, calling discoverUsableModels(store) returns one entry per provider with valid credentials plus their live model lists.

License

MIT — see LICENSE.