@chirag127/ai

One AI interface, two backends. Same chat(messages, opts) -> Promise<string> contract both sides.

• ./server — NVIDIA NIM (OpenAI-compatible). Key-safe: reads NVIDIA_API_KEY from env. Server-side only, never ship to client.
• ./client — Puter.js. Keyless: users auth via Puter, no key leaves the browser. :free models only.

Framework-agnostic, zero deps.

Install

npm i @chirag127/ai

Server — NVIDIA NIM (key-safe)

Key from env NVIDIA_API_KEY (GH Secret). Never embedded, never sent to a browser. Tries 3 models in order.

import { chat } from '@chirag127/ai/server';

// NVIDIA_API_KEY in env (or pass opts.apiKey)
const reply = await chat([{ role: 'user', content: 'Explain HMAC in one line.' }]);
console.log(reply);

// pick a model (still falls back through the rest on error)
await chat('hi', { model: 'meta/llama-3.3-70b-instruct' });

Throws if NVIDIA_API_KEY is missing. Models tried in order: nvidia/llama-3.3-nemotron-super-49b-v1 → meta/llama-3.3-70b-instruct → deepseek-ai/deepseek-r1.

Client — Puter.js (keyless)

Loads js.puter.com/v2 on first call. No key anywhere; the user authorizes via Puter.

import { chat, listFreeModels } from '@chirag127/ai/client';

const reply = await chat('Write a haiku about caching.');
console.log(reply);

// live :free catalog
console.log(await listFreeModels());

:free models tried in order: nvidia/nemotron-3-super-120b-a12b:free → openai/gpt-oss-20b:free → google/gemma-4-31b-it:free.

Shared

import { MODELS, toMessages } from '@chirag127/ai';
MODELS.server; // string[]
MODELS.client; // string[] (:free)

Both backends accept a string prompt or a messages array ({ role, content }[]).

License

MIT