@tozil/core

Core engine for Tozil — know what every AI user costs you.

Handles event buffering, context propagation, and the plugin system. Provider-specific instrumentation is in separate packages.

Install

npm install @tozil/core

Usage

import { init } from "@tozil/core";
import { anthropic } from "@tozil/anthropic";
import { openai } from "@tozil/openai";

init({
  apiKey: "tz_...", // or set TOZIL_API_KEY env var
  instrumentations: [anthropic(), openai()],
});

Every AI call in your app is now tracked. No code changes needed beyond this.

What it does

• Buffers usage events in memory, flushes every 5 seconds
• Propagates user/endpoint context via AsyncLocalStorage
• Provides a plugin interface (Instrumentation) for provider packages
• Never throws — if tracking fails, your app keeps running
• Sends only metadata (model, tokens, latency) — no prompts or completions

API

init(options)

| Option | Type | Default | Description | |--------|------|---------|-------------| | apiKey | string | process.env.TOZIL_API_KEY | Your Tozil API key | | baseUrl | string | https://app.tozil.dev/api/v1 | API endpoint | | instrumentations | Instrumentation[] | [] | Provider plugins to activate | | flushInterval | number | 5000 | Flush interval in ms | | maxBatchSize | number | 100 | Max events per flush | | debug | boolean | false | Log tracking events to console |

withContext(ctx, fn)

Run a function with user/endpoint context:

import { withContext } from "@tozil/core";

withContext({ userId: "user_123", endpoint: "/api/chat" }, async () => {
  // AI calls in here are tagged with this user and endpoint
  await anthropic.messages.create({ ... });
});

shutdown()

Flush remaining events and clean up. Call on process exit.

Provider packages

| Package | Provider | |---------|----------| | @tozil/anthropic | Anthropic (Claude) | | @tozil/openai | OpenAI (GPT, o1) | | @tozil/google-ai | Google Generative AI (Gemini) | | @tozil/vercel-ai | Vercel AI SDK | | @tozil/langchain | LangChain / LangGraph |

License

MIT