@auix/prism

LLM tracing SDK for auix — collect LLM traces with zero friction.

Supports Vercel AI SDK, OpenAI, Anthropic, and manual tracing.

Install

npm install @auix/prism

Peer dependencies (install only what you use):

• ai (>=6.0.0) + @ai-sdk/provider (>=2.0.0) — for AI SDK integration
• openai (>=4.0.0) — for OpenAI integration
• @anthropic-ai/sdk (>=0.30.0) — for Anthropic integration

Quick Start

AI SDK (recommended)

prismAISDK wraps any AI SDK model and automatically captures traces, token usage, latency, and tool calls. It creates a root trace internally and returns { model, end }:

import { AuixPrism, prismAISDK } from "@auix/prism";
import { openai } from "@ai-sdk/openai";
import { generateText } from "ai";

const tracer = new AuixPrism({ apiKey: "your-api-key" });

const traced = prismAISDK(tracer, openai("gpt-4o"), {
  name: "my-chat",
  tags: ["production"],
  endUserId: "[email protected]",
  metadata: { threadId: "t_123" },
});

const { text, usage } = await generateText({
  model: traced.model,
  prompt: "Hello!",
});

// Token usage is captured automatically from spans — just call end():
await traced.end();

// Or pass explicit values to override:
await traced.end({
  output: text,
  totalTokens: usage.totalTokens,
  inputTokens: usage.inputTokens,
  outputTokens: usage.outputTokens,
});

Multi-step tool use is handled correctly — each LLM call becomes a child span under the root trace, and token usage is automatically aggregated:

import { streamText, stepCountIs } from "ai";

const traced = prismAISDK(tracer, openai("gpt-4o"), {
  name: "agent",
  tags: ["chat"],
});

const result = streamText({
  model: traced.model,
  prompt: "What's the weather?",
  tools: { /* ... */ },
  stopWhen: stepCountIs(6),
});

for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}

await traced.end();

OpenAI

prismOpenAI returns a proxied client that traces chat.completions.create calls:

import { AuixPrism, prismOpenAI } from "@auix/prism";
import OpenAI from "openai";

const tracer = new AuixPrism({ apiKey: "your-api-key" });
const client = prismOpenAI(tracer, new OpenAI(), {
  name: "my-chat",
  tags: ["production"],
});

const response = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Hello!" }],
});

Anthropic

prismAnthropic returns a proxied client that traces messages.create calls:

import { AuixPrism, prismAnthropic } from "@auix/prism";
import Anthropic from "@anthropic-ai/sdk";

const tracer = new AuixPrism({ apiKey: "your-api-key" });
const client = prismAnthropic(tracer, new Anthropic(), {
  name: "my-chat",
  tags: ["production"],
});

const response = await client.messages.create({
  model: "claude-sonnet-4-20250514",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Hello!" }],
});

Manual Tracing

For full control over trace structure:

import { AuixPrism } from "@auix/prism";

const tracer = new AuixPrism({ apiKey: "your-api-key" });

const trace = tracer.startTrace({
  name: "rag-pipeline",
  metadata: { userId: "u_123" },
});

const span = trace.startSpan({
  name: "vector-search",
  type: "retrieval",
  input: { query: "How do I reset my password?" },
});

span.end({ output: { results: 5 }, status: "completed" });

trace.end({ output: "Here's how to reset...", status: "completed" });

await tracer.destroy();

API

new AuixPrism(config)

| Option | Type | Default | Description | |--------|------|---------|-------------| | apiKey | string | required | Your auix API key | | baseUrl | string | "https://api.auix.dev" | API endpoint | | sessionId | string | — | Group traces under a session | | transport | (events) => Promise<void> | — | Custom transport (bypasses HTTP) | | onFlushError | (error) => void | — | Called when flush fails after retry |

prismAISDK(tracer, model, options?)

Returns { model, end }. The model is a wrapped AI SDK model. Call end(opts?) after the generation completes to finalize the root trace.

| Option | Type | Description | |--------|------|-------------| | name | string | Trace name (defaults to model ID) | | tags | string[] | Tags for filtering | | metadata | Record<string, unknown> | Arbitrary metadata | | endUserId | string | End user identifier |

prismOpenAI(tracer, client, options?)

Returns a proxied OpenAI client. Same options as prismAISDK plus parentTraceId for nesting under an existing trace.

prismAnthropic(tracer, client, options?)

Returns a proxied Anthropic client. Same options as prismOpenAI.

tracer.startTrace(options)

Returns a TraceHandle for manual tracing.

trace.startSpan(options)

Creates a child span. type can be "llm", "tool", "retrieval", or "custom".

tracer.destroy()

Flushes pending events and stops the background timer. Call this when done tracing.

License

Proprietary — see LICENSE.