Real transcript · nothing staged

We gave Claude a TraceBug report. Here's exactly what happened.

A bug was triggered on our live sandbox, captured with the real widget, exported to one .html file — and an AI coding agent read it through TraceBug's MCP server. Five tool calls from crash to root cause. The outputs below are unedited (trimmed for length only).

The setup. The live sandbox has two intentional bugs: applying the coupon SAVE20 throws a TypeError, and placing the order calls an endpoint that doesn't exist. We captured both with two clicks, exported the report, and registered the MCP server:

claude mcp add tracebug -- npx -y tracebug mcp

Then we asked the agent to investigate. No other context was given.

1list_bug_reports()

The agent starts blind — it only knows a report exists somewhere.

{
  "reports": [{
    "file": "tracebug-replay-93e6615d-2026-07-19T08-07-00.html",
    "title": "Try.html Page: \"Place order\" Action Fails — TypeError",
    "summary": "API POST /api/orders failed with 404 when clicking
                'Place order' button on /try.html",
    "severity": "critical",
    "rootCause": {
      "hint": "API POST /api/orders failed with 404 after clicking 'Place order'",
      "confidence": "high"
    },
    "counts": { "consoleErrors": 1, "networkFailures": 2, "userActions": 6 }
  }],
  "nextStep": "Call get_bug_report on a file for the full overview..."
}

2get_bug_report("tracebug-replay-93e6615d")

The report answers with an investigation guide — it tells the agent what to look at, in priority order.

{
  "title": "Try.html Page: \"Place order\" Action Fails — TypeError",
  "environment": "Chrome 150.0.0.0 · Windows 10/11 · 1280x800",
  "sessionDurationMs": 463,
  "investigationGuide": [
    "[HIGH] get_network_activity — 2 failed requests captured...",
    "[HIGH] get_console_errors — 1 console error captured; stack traces
            point at the failing file and function.",
    "[MEDIUM] get_repro_steps — the plain-English steps and session
            timeline show what the user did leading up to the bug.",
    "Finally: cross-reference the findings with the codebase..."
  ]
}

3get_console_errors("tracebug-replay-93e6615d")

The stack trace names the exact file and line.

{
  "count": 1,
  "entries": [{
    "level": "error",
    "message": "Uncaught TypeError: Cannot read properties of undefined
                (reading 'discount')",
    "stack": "TypeError: Cannot read properties of undefined (reading 'discount')
    at http://localhost:3799/try.html:140:31"
  }]
}

4get_network_activity("tracebug-replay-93e6615d")

The second failure, with method, status, and timing.

{
  "failureCount": 2,
  "requests": [
    { "method": "POST", "url": "/api/orders", "status": 404, "durationMs": 12 },
    { "method": "GET",  "url": "http://localhost:3799/api/orders", "status": 404 }
  ]
}

5get_repro_steps("tracebug-replay-93e6615d")

A millisecond-resolution timeline of what the user actually did.

"timeline": [
  { "elapsed": "00:00.00", "description": "click \"Apply coupon\"" },
  { "elapsed": "00:00.06", "description": "Uncaught TypeError: Cannot read
                            properties of undefined (reading 'discount')" },
  { "elapsed": "00:00.27", "description": "input \"coupon\" = \"SAVE20\"" },
  { "elapsed": "00:00.44", "description": "click \"Place order\"" },
  { "elapsed": "00:00.46", "description": "POST /api/orders → 404 (12ms)" }
]

The agent's diagnosis — from the report alone

Bug 1 — the TypeError. The stack trace points at try.html:140. That line reads coupons[code].discount — but the coupon map only defines WELCOME10, so entering SAVE20 reads .discount off undefined. Fix: guard unknown codes before the read.

Bug 2 — the failed order. POST /api/orders → 404 in 12ms: the endpoint doesn't exist on this host. The timeline proves the order in which the user hit both bugs — 463 milliseconds, fully reconstructed.

Total investigation: 5 tool calls. Nobody explained the bug, shared a screen, or answered "what were you doing when it broke?"

Run this exact experiment yourself

The sandbox, the bugs, and the MCP server are all public. Capture, export, and hand the report to your own agent — it takes about two minutes.

Open the live sandbox MCP setup guide

Transcript generated 2026-07-19 by Claude reading a real TraceBug export via npx -y tracebug mcp. Outputs trimmed for length, never reworded.