@botzone/cost-sdk

Cost-tracking SDK for Anthropic, OpenAI, and Gemini clients. Wraps your existing LLM client and ships per-call usage to your Cost dashboard. Adds zero measurable latency to the host call.

Install

npm install @botzone/cost-sdk

Usage

import Anthropic from "@anthropic-ai/sdk";
import { wrap } from "@botzone/cost-sdk";

const anthropic = wrap(new Anthropic(), {
  apiKey: process.env.COST_API_KEY,
  route: "follow-up-draft",
  enabled: process.env.COST_ENABLED === "true",
});

wrap auto-detects the provider. Same surface for OpenAI and Gemini:

import OpenAI from "openai";
import { GoogleGenerativeAI } from "@google/generative-ai";

const openai = wrap(new OpenAI(), { route: "summariser" });
const gemini = wrap(new GoogleGenerativeAI(key), { route: "classifier" });

Options

| field | meaning | | -------------- | ----------------------------------------------------------- | | apiKey | Cost API key. Defaults to process.env.COST_API_KEY. | | endpoint | Ingestion URL. Defaults to https://cost.botzone.ai. | | route | Logical name shown in the dashboard (strongly recommended). | | userId | Optional end-user id: sha256-hashed before send. | | featureTag | Free-form label for slicing. | | enabled | Master switch. false = passthrough. | | captureBodies| Capture full prompt + response (default false). Set true per route to enable verify-downgrade replay. |

What gets captured

By default (metadata only): token counts (including Anthropic prompt-cache reads / writes and OpenAI cached prompt tokens), latency, model, route, user id (hashed in the SDK before send), feature tag, SHA-256 hashes of prompt and response. Computed USD cost is added server-side from the live pricing table.

Opt-in (captureBodies: true): the raw request and response JSON are sent in addition to the metadata above. This is required for the verify-downgrade feature, which replays a sample of your real traffic on a cheaper model. Set it per-route so only the traffic you want evaluated has its bodies sent.

Cost stores raw bodies for 30 days on the free tier, 90 days on paid plans, then purges them automatically. End-user identifiers you pass via userId are hashed with SHA-256 in the SDK; the plaintext never leaves your process.