Monitor your Vercel AI SDK features in production.

Install

npm install @agentping/sdk

Instrument

import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
import * as agentping from "@agentping/sdk";

agentping.init({ apiKey: process.env.AGENTPING_API_KEY });

// One run per call, attributed to the customer and feature.
const run = agentping.run("copilot", { customerId: "acme-corp", feature: "summarise" });

const { text } = await generateText({
  model: openai("gpt-4o-mini"),
  prompt: "…",
  ...agentping.withAgentPing(run),
});

await run.finish({ status: "success" });

Full setup, environment variables and options are in the Vercel AI SDK docs.

• Cost per run, attributed to the customer and feature you tag.
• Run status and latency for every AI call in the product.
• Cost spike alerts when a feature breaks its baseline.
• Output quality checks so a regression is caught early.
• Streaming-aware: a streamed response is still one clean run record.

Every call, with its cost

Each generateText or streamText call lands as a run with the provider and model taken from the AI SDK's response, input and output tokens summed across every step, cached and reasoning tokens when reported, the finish reason and the latency. Agent Ping prices it from a maintained rate card, so the cost column is there without any maths in your code.

The tools the model used

Each tool the model called appears on the run with its arguments and result, and a tool that errored is marked as an error. Turn on perStep and a tool-using call shows each model round trip separately, so you can see which step was the expensive one.

Spend by feature and by customer

Tag the run with customerId and feature and the cost views break down that way: which AI feature in your product costs the most, which accounts drive it, and how both are trending against their baseline.

Quality and failures next to cost

With scoring on, a sampled share of runs is judged against your rubric, so a weaker answer shows up as a falling line next to the cost line. A call that throws before onFinish never reaches the helper, so finish the run with an error status in your catch block and failures show up beside the rest.

• A run list for Vercel AI SDK, each row showing cost, status and latency, filterable by agent, customer and feature.
• Cost by agent, customer and feature, so you can see which accounts and features drive AI spend, not just one provider total.
• An incident the moment a scheduled run misses its window or a cost spike breaks its baseline, delivered to Slack, email, PagerDuty or a webhook.
• A quality trend once you turn on scoring, so a regression in your Vercel AI SDK output shows up as a falling line, not a support ticket.