MLflow Typescript SDK - Vercel AI

Seamlessly integrate MLflow Tracing with Vercel AI SDK to automatically trace your AI API calls.

| Package | NPM | Description | | -------------------- | --------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------- | | @mlflow/vercel | | Auto-instrumentation integration for Vercel AI SDK. |

Installation

npm install @mlflow/vercel

The package includes @opentelemetry/api and @opentelemetry/sdk-trace-base as peer dependencies. Depending on your package manager, you may need to install them separately.

Quickstart

Start MLflow Tracking Server. If you have a local Python environment, you can run the following command:

pip install mlflow
mlflow server --port 5000

If you don't have Python environment locally, MLflow also supports Docker deployment or managed services. See Self-Hosting Guide for getting started.

Set up the MLflow span processor and use the Vercel AI SDK with telemetry enabled:

import { MLflowSpanProcessor } from '@mlflow/vercel';
import { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-proto';
import { NodeTracerProvider } from '@opentelemetry/sdk-trace-node';
import { generateText } from 'ai';
import { openai } from '@ai-sdk/openai';

const provider = new NodeTracerProvider({
  spanProcessors: [
    new MLflowSpanProcessor(
      new OTLPTraceExporter({
        url: 'http://localhost:5000/api/2.0/otel/v1/traces',
        headers: {
          'x-mlflow-experiment-id': '<your-experiment-id>',
        },
      }),
    ),
  ],
});
provider.register();

const result = await generateText({
  model: openai('gpt-5'),
  prompt: "What's the weather like in Seattle?",
  experimental_telemetry: { isEnabled: true },
});

Databricks

To send traces to a Databricks Unity Catalog table, set the OTLP exporter URL to <DATABRICKS_HOST>/api/2.0/otel/v1/traces and include the following headers:

• Authorization: Bearer <your-databricks-token>
• X-Databricks-UC-Table-Name: <catalog>.<schema>.<table_prefix>_otel_spans

Note: Do not set the x-mlflow-experiment-id header when using Databricks.

Attribute Translation

The Vercel AI SDK emits spans with ai.* attributes. MLflowSpanProcessor translates these into MLflow's format:

| Vercel AI SDK | MLflow | Description | | -------------------------------------------- | ------------------------------------------ | -------------------------------- | | ai.operationId | mlflow.spanType | Span type (LLM, TOOL, EMBEDDING) | | ai.prompt.* / ai.response.* | mlflow.spanInputs / mlflow.spanOutputs | Structured request/response data | | ai.model.id | mlflow.llm.model | Model name | | ai.model.provider | mlflow.llm.provider | Provider name | | ai.usage.promptTokens / completionTokens | mlflow.chat.tokenUsage | Token usage for cost tracking | | (chat spans) | mlflow.message.format = "vercel_ai" | Enables chat UI rendering |

Documentation

• MLflow Tracing
• Vercel AI SDK Telemetry
• Databricks OTEL Collector

License

This project is licensed under the Apache License 2.0.