A zero-dependency tracer + local dashboard for LangChain (and other) agents. When an agent misbehaves — calls the wrong tool, hangs, or silently fails — StepGlass shows you the full run as a timeline: every tool call, every LLM call, what went in, what came out, and exactly where it broke.

StepGlass dashboard demo

Why

Agent frameworks are good at making agents run. They're not good at showing you what happened when a run goes wrong. Most teams end up grepping through console logs trying to reconstruct a call sequence after the fact.

StepGlass does one thing: it records every step of an agent run to a local file, and gives you a visual timeline to inspect it. No cloud account, no API key, no data leaving your machine.

Install

npm install stepglass

Quick start

Generate a sample trace and see the dashboard (no API keys needed):

npx stepglass dashboard

Using it with a real agent

import { createTraceHandler } from "stepglass";

const { handler, logger } = createTraceHandler({ label: "support-bot run" });

const result = await agentExecutor.invoke(
  { input: userMessage },
  { callbacks: [handler] }
);

logger.finish("completed");

Then run:

npx stepglass dashboard

This opens a local dashboard at http://localhost:4550 showing every run recorded in .stepglass/. Click any step in the timeline to see its full input, output, or error.

Not using LangChain?

The core TraceLogger is framework-agnostic — call start() / end() / error() around any function:

import { TraceLogger } from "stepglass";

const logger = new TraceLogger({ label: "my custom agent" });

const step = logger.start("tool_start", "fetch_weather", { city: "London" });
try {
  const result = await fetchWeather("London");
  logger.end("tool_end", step, "fetch_weather", result);
} catch (err) {
  logger.error("tool_error", step, "fetch_weather", err);
}

logger.finish("completed");