@anyway-sh/node-server-sdk

OpenTelemetry-based observability SDK for LLM applications. Automatically instruments LLM provider calls, tracks token usage, and calculates costs.

Installation

npm install @anyway-sh/node-server-sdk

Quick Start

import { initialize } from "@anyway-sh/node-server-sdk";

initialize({
  apiKey: process.env.ANYWAY_API_KEY,
  appName: "my-app",
});

ESM / Next.js

ESM projects must pass provider modules explicitly since OpenTelemetry's require-based instrumentation doesn't work with ESM imports:

import { initialize } from "@anyway-sh/node-server-sdk";
import * as OpenAIModule from "openai";

initialize({
  apiKey: process.env.ANYWAY_API_KEY,
  appName: "my-app",
  instrumentModules: {
    openAI: OpenAIModule.OpenAI,
  },
});

Tracing Workflows

Use withWorkflow, withTask, withAgent, and withTool to create structured trace hierarchies:

import { withWorkflow, withTask } from "@anyway-sh/node-server-sdk";

async function main() {
  await withWorkflow({ name: "my-pipeline" }, async () => {
    const result = await withTask({ name: "generate-response" }, async () => {
      return openai.chat.completions.create({
        model: "gpt-4",
        messages: [{ role: "user", content: "Hello" }],
      });
    });
    return result;
  });
}

Class Decorators

import { workflow, task } from "@anyway-sh/node-server-sdk";

class MyPipeline {
  @workflow({ name: "run" })
  async run() {
    return this.step();
  }

  @task({ name: "step" })
  async step() {
    // ...
  }
}

Association Properties

Attach metadata to traces for filtering and grouping:

await withWorkflow(
  {
    name: "chat",
    associationProperties: { userId: "user-123", sessionId: "abc" },
  },
  async () => {
    // All spans in this workflow will carry these properties
  },
);

Configuration

Key options for initialize():

| Option | Default | Description | |--------|---------|-------------| | apiKey | ANYWAY_API_KEY env var | API key for sending traces | | appName | npm_package_name | Application name in traces | | baseUrl | ANYWAY_BASE_URL or https://api.traceloop.com | Trace collector endpoint | | disableBatch | false | Send spans immediately (for local dev) | | exporter | OTLP exporter | Custom OpenTelemetry SpanExporter | | processor | BatchSpanProcessor | Custom OpenTelemetry SpanProcessor | | instrumentModules | — | Explicit module references for ESM projects | | pricingEnabled | true | Calculate and attach cost attributes to spans | | pricingJsonPath | — | Path to custom pricing JSON file | | tracingEnabled | true | Enable/disable tracing entirely | | silenceInitializationMessage | false | Suppress startup console message |

Supported Providers

LLM providers and vector databases instrumented automatically:

• OpenAI — chat completions, embeddings
• Anthropic — messages
• AWS Bedrock — invoke model
• Google Vertex AI — generative models
• Cohere — chat, embed, rerank
• Together AI — chat completions
• LangChain — chains, agents, tools
• LlamaIndex — query engines, retrievers
• Pinecone — vector operations
• ChromaDB — collections, queries
• Qdrant — points, search
• MCP — Model Context Protocol client calls

Pricing

Cost calculation is enabled by default. The SDK matches model names from spans against bundled pricing data and sets these attributes:

• gen_ai.usage.input_cost
• gen_ai.usage.output_cost
• gen_ai.usage.cost

To use custom pricing data:

initialize({
  pricingJsonPath: "./my-pricing.json",
});

To disable:

initialize({
  pricingEnabled: false,
});

License

Apache-2.0