@virajmishra1/bench-mastra

Auto-instrument Mastra agents with Bench. One line — every .generate() becomes a Bench task, every tool execution becomes a subtask.

npm i @virajmishra1/bench-mastra @virajmishra1/bench-sdk

Quickstart

import { Agent } from "@mastra/core";
import { observe } from "@virajmishra1/bench-sdk";
import { wrapAgent } from "@virajmishra1/bench-mastra";

const bench = observe({
  apiKey: process.env.BENCH_KEY!,
  agent: "my-mastra-agent",
});

const researcher = wrapAgent(
  new Agent({
    name: "researcher",
    instructions: "You research things.",
    model: { provider: "OPEN_AI", name: "gpt-4o" },
    tools: { search: searchTool, fetch: fetchTool },
  }),
  bench,
);

// Every call shows up live in your Bench dashboard.
const result = await researcher.generate("research the OTLP spec");

Pricing

Mastra doesn't bundle per-model pricing. Pass pricePerMTok to capture cost:

wrapAgent(myAgent, bench, {
  pricePerMTok: { input: 2.5, output: 10 }, // gpt-4o per 1M tokens
});

Wrap a tool directly

import { wrapTool } from "@virajmishra1/bench-mastra";

const tracedSearch = wrapTool(searchTool, bench, "google-search");

License

MIT