@spendgraph/sdk

Track LLM input/output tokens and cost. Three functions, zero dependencies, fail-open.

npm install @spendgraph/sdk

import Anthropic from "@anthropic-ai/sdk";
import { SpendGraph } from "@spendgraph/sdk";

const meter = new SpendGraph({ apiKey: process.env.SPENDGRAPH_API_KEY, baseUrl });
const anthropic = meter.wrap(new Anthropic());

Use the client exactly as before. Usage is read off each reply and reported for you, and spend appears on the dashboard within seconds.

What you get

• One line to wire in. wrap() proxies an Anthropic or OpenAI client and reads the provider's own usage block, cache counts included.
• It never throws into your code. Batched, non-blocking, and it drops data rather than take your app down when the meter is having a bad day.
• Nothing about the conversation is sent — the model, the counts, and whatever metadata you attached.
• The whole API, typed. Spendgraph reaches prompts, tools, stats and events, with retries and error shapes decided once.

Docs

| | | | --- | --- | | Quickstart | Two minutes to your first tracked call | | Tracking | wrap, track and flush | | The client | Spendgraph, and the two halves it reaches | | Errors and retries | What throws, what retries, what never will | | HTTP API | Ingest, stats, events, CSV export | | Integrations | LangChain, LangGraph, Vercel AI | | Budgets and alerts | Thresholds, webhooks, what never happens |

License

Apache-2.0