@runbit/devtools

Zero-instrumentation observability. The runtime already emits every event — devtools folds them into per-run traces.

npm install @runbit/devtools

import { runtime } from "@runbit/runtime";
import { devtools } from "@runbit/devtools";

const app = runtime();
app.use(devtools());   // that's the whole setup

await app.run(assistant, input);

Prints the full timeline when a run ends:

Trace 7323a0e7  (ok, 1240ms)
  +0ms      runtime:start
  +2ms      agent:start  "Weather in Tokyo?"
  +5ms      llm:request  model=gpt-4o-mini
  +820ms    llm:response  model=gpt-4o-mini tokens=120→30
  +825ms    tool:start  weather
  +826ms    tool:end  weather
  +1240ms   runtime:end
  ── 2 LLM · 1 tool · 355 tokens · $0.0001

API

• devtools({ log, price }) — a plugin; price(model, usage) computes cost.
• .traces() / .trace(runId) / .last() — read traces programmatically.
• .print(runId?) — print a timeline.
• .onTrace(cb) — hook each finished run (send to OpenTelemetry, Langfuse, a UI, …).

No tracing SDK, no external service. The runtime knew everything already.

Part of Runbit — the AI Runtime. Everything executes through one small, event-driven runtime; agents, workflows, RAG, and observability are packages on top.

• Umbrella: @runbit/runtime
• License: MIT