OpenTelemetry

lanjakToken Usage & Limits
Get it for macOS

Send per-turn token, cost and quota data from bb to the OpenTelemetry collector you already run. Each finished turn becomes a span with the model, input, output and cache tokens, and cost when the provider reports it; child threads join the trace of their root thread. Histograms and counters track token usage, turn duration, cost, turns and thread transitions, and gauges follow provider quota windows (used percent, reset time). Prompts, replies, tool text, thread titles, file paths and error messages are never exported, and no setting turns that on. Needs an OpenTelemetry collector that you run; nothing is sent until you set its OTLP/HTTP endpoint.

Overview

View source

What you get

  • One span for each finished turn, with model, token counts, cache tokens and cost, in GenAI semantic-convention attributes.
  • Thread, project and environment ids, plus the project name, on every span. A delegated thread joins the trace of its parent.
  • Token, cost, duration and thread-lifecycle metrics.
  • Provider quota windows (used percent, reset time) as gauges.

Privacy

The plugin never exports prompts, replies, tool input or output, thread titles, file paths or error messages. There is no setting to turn that on.

Requirements

An OpenTelemetry collector that you run and that accepts OTLP/HTTP.

Set up

Set the OTLP/HTTP endpoint of your collector, and the headers it needs. Nothing is sent until the endpoint is set.

More from lanjak

RouteHand mechanical edits to a cheaper agent, ask another model, or get a ranked decision without leaving the turn. route_worker runs a task on any BB harness and model you list, in the same environment, and returns its final message. route_ask sends one prompt to any OpenAI-compatible API, and route_decide asks a System One decision model such as TypeSafe Jev for a pick with probabilities. Each tool needs its own provider or endpoint and key, billed by that service.New

More in Token Usage & Limits 31

Understand or control token, context-window, and provider-quota use.

View all