FOR AI ENGINEERS

Read agent traces
at the speed of ⌘K.

Tool calls, spans, structured outputs, eval sets — it's all deeply-nested JSON now. Ask which span failed, how many tokens a run burned, or which evals regressed. QDev writes the jq and runs it locally.

✓Triage a failing run without jq gymnastics or a scratch script
✓Pass rates & token counts are exact — jq computes them, not a model
✓Prompts & completions stay local — nothing uploaded

Add to Chrome Add to Firefox

⌘K

try

● TRACE.JSON — RUN a3f1 · 214 SPANS · LOCAL

{
  "spans": [
    {
      "name": "plan",
      "kind": "llm",
      "model": "gpt-5",
      "ms": 1840,
      "tokens": 2110,
      "status": "ok"
    },
    {
      "name": "search_web",
      "kind": "tool",
      "model": null,
      "ms": 920,
      "tokens": 0,
      "status": "ok"
    },
    {
      "name": "fetch_page",
      "kind": "tool",
      "model": null,
      "ms": 210,
      "tokens": 0,
      "status": "error"
    }
  ]
}

↑ Ask about the run — jq scans all 214 spans locally. Only the schema is ever sent, never prompt or completion text.

exact metrics · reproducible jq · nothing uploaded

SOUND FAMILIAR?

Debugging the black box

Malformed tool-call JSON

The model returned arguments that don't parse, and the failure is buried three levels into a 200-span trace.

Eval sets too big to open

A 60 MB JSONL of results won't load in your editor, and you just need the pass rate per suite.

jq you rewrite every run

group_by, select on a nested status, sum the tokens — you know the query exists, you just don't want to author it again.

ASK THE TRACE ANYTHING

From raw spans to a root cause

🧯

Triage failures

Jump straight to the span that errored, or every tool call whose output failed to parse.

spans where status is error, with the message

⏱️

Latency & cost

Sum tokens by model, find the slowest step, total the wall-clock of a run — exact, via jq.

total tokens and cost by model

🧪

Score eval runs

Pass rate per suite, regressions vs a baseline, the cases that flipped — no scratch notebook.

pass rate grouped by suite

🔧

Inspect tool calls

Pull every function call's name and arguments across the conversation into one flat list.

list every tool call name and its arguments

FITS YOUR EVAL LOOP

Every format you already work in

TRACES

Agent spans

OpenTelemetry-style or your own — deeply nested runs, flattened by one ask.

TOOL CALLS

Structured output

Function-calling args and JSON-mode responses, validated against the shape.

JSONL

Eval sets

Line-delimited results — aggregate pass/fail without a parsing script.

LOGS

Request logs

Prompt/response logs and streaming NDJSON, grouped and counted locally.

# QDev: "eval cases that regressed vs baseline" → editable jq [.cases[] | select(.pass == false)] | group_by(.suite) | map({suite: .[0].suite, failed: length}) | sort_by(-.failed)

● YOUR PROMPTS ARE YOURS

Traces never leave the tab.

Prompts, completions, tool outputs, customer data caught in a span — the values stay local. Only the schema (span shape & field types) and your question are ever sent, and only on the hosted tier. Want zero egress? Run local AI, free.

✓jq runs in a sandboxed Web Worker — no network access

✓Hosted proxy is stateless & zero-retention — logs nothing

✓No trace exporter, no vendor account, no data pipeline to set up

Your agent failed. Find out why.

Install free, keep your traces local, and let ⌘K write the jq that finds the failing span.