llm-intel

Model intelligence for every LLM.

llm-intel sources model metadata and pricing from OpenRouter, so you can look up capabilities and calculate token costs without maintaining your own data tables.

Features

• Look up any model's capabilities, context window, and pricing by ID
• Calculate token costs with exact precision (powered by bignumber.js)
• Two purpose-built APIs: a server client (fetches from OpenRouter) and a browser client (reads pre-fetched JSON)
• Three-tier caching: memory -> disk -> network
• Full TypeScript support with generated ModelId types

View Demo →

Installation

# npm
npm install @basisoasis/llm-intel

# pnpm
pnpm add @basisoasis/llm-intel

# yarn
yarn add @basisoasis/llm-intel

# bun
bun add @basisoasis/llm-intel

Usage

Server (LLMIntel)

Use this in Node.js / server-side environments. It fetches model data from OpenRouter, with disk and memory caching built in.

import { LLMIntel } from "@basisoasis/llm-intel";

// Instantiate a provider client
const client = await LLMIntel.create({ provider: 'openrouter' });

// Resolve a model by ID
const model = await client.getModel(
  'anthropic/claude-4.6-sonnet-20260217'
);

if (!model) throw new Error('Model not found!');

const cost = client.calculateCost(model, {
  inputTokens: 20_000,
  outputTokens: 1700,
});

console.log(client.formatCostResult(cost));
/* {
  inputCost: "$0.06",
  outputCost: "$0.03",
  cacheReadCost: null,
  cacheWriteCost: null,
  imageCost: null,
  requestCost: null,
  totalCost: "$0.09",
  currency: "USD",
  warnings: [],
} */

Standalone function

For one-off lookups without instantiating a client:

import { getModelInfo } from "@basisoasis/llm-intel";

const result = await getModelInfo("anthropic/claude-3-5-sonnet", {
  provider: "openrouter",
  apiKey: process.env.OPENROUTER_API_KEY,
});

Browser / SPA (LLMIntelClient)

Use this when you already have the model JSON (e.g. fetched server-side and passed to a SPA, or bundled at build time). No API key required.

import { LLMIntelClient } from "@basisoasis/llm-intel/client";

// Hydrate from a URL your server exposes
const client = new LLMIntelClient({
  models: "/api/models", // returns ModelsResult JSON
  cacheTtl: 5 * 60 * 1000, // 5 minutes
});

// Or hydrate statically from a pre-loaded array
const client = new LLMIntelClient({ models: modelDataArray });

const model = await client.getModel("google/gemini-2.5-pro");
if (!model) throw new Error('Model not found!');

const cost = client.calculateCost(model, {
  inputTokens: 2000,
  outputTokens: 800,
});

console.log(client.formatCost(cost.inputCost));  // $0.0025
console.log(client.formatCost(cost.outputCost)); // $0.008
console.log(client.formatCost(cost.totalCost));  // $0.01

API Reference

LLMIntel (server)

LLMIntelClient (browser)

Same getModel, getModels, calculateCost, formatCost, and formatCostResult methods. Takes either a URL or a pre-loaded ModelData[] array.

getModelInfo(modelId, opts) (standalone)

Fetches a single model without creating a client. Useful for serverless functions or scripts.

Caching

LLMIntel uses a three-tier cache:

1. Memory: fastest; per-instance, invalidated by TTL
2. Disk: survives process restarts
3. Network: fetches fresh data from OpenRouter

Configure the TTL via cacheTtl in milliseconds (default: 86_400_000 — 24 hours).

LLMIntelClient uses memory caching only (no disk access in the browser).

Configuration

All options are optional — the library falls back to environment variables and then built-in defaults.

LLMIntel.create({
  provider: "openrouter",
  openRouterApiKey: process.env.LLM_INTEL_OPEN_ROUTER_API_KEY,
  cacheTtl: 86_400_000,
  cacheDir: ".cache",
});

Providers

Currently, OpenRouter is the only supported provider. The library has been designed with a provider abstraction layer, so support for additional data sources can be added in the future without breaking changes to the public API.

License

MIT