agent-otel-bridge mark agent-otel-bridgeThe Live Dashboard · v0.6.0
How agent-otel-bridge works

Fast on the developer path. Rich in your monitoring stack.

agent-otel-bridge separates the tiny hook that runs inside your coding CLI from the background worker that enriches, redacts, and exports standard OpenTelemetry signals.

Step 01 Lightweight CLI Hook

Instant handoff, fail-open by design

When Claude Code, Codex, or Gemini CLI starts a turn or runs a tool, a tiny native hook forwards the event locally in under a millisecond and exits immediately—never waiting on network I/O.

Step 02 Intent & Privacy Filter

Classify behavior and scrub secrets

A background worker groups tool calls into meaningful engineering categories (reading, editing, testing, Git, network), links parent and sub-agent traces, and redacts credentials and raw prompt bodies.

Step 03 Standard OTLP Export

Drop into Grafana, Datadog, or Honeycomb

Traces and metrics flow over standard OpenTelemetry (OTLP) to your existing collector—complete with prebuilt dashboards for adoption, tool reliability, and subscription rate-limit headroom.

Questions answered & boundaries

What you can see live—and what stays private.

Observability for engineering tools should empower teams to spot friction and risk without turning into keystroke surveillance.

Operational question What agent-otel-bridge shows you Privacy & safety boundary
How are teams adopting AI coding tools? Active sessions, turns, and token usage broken down by tool (Claude Code, Codex, Gemini CLI) and model. Prompt text and source code bodies are excluded by default.
What kinds of actions are agents taking? Behavioral breakdown across code reading, file editing, test execution, Git operations, and external network calls. API keys, bearer tokens, and environment secrets in command strings are automatically scrubbed.
Are we about to hit subscription rate limits? Live quota utilization and remaining headroom gauges so teams know before a rate-limit wall interrupts delivery. Never proxies or intercepts your model API traffic—reads only local hook metadata.
What happens if the OTLP collector goes down? The local hook fails open in milliseconds so developer terminals never hang or error out. Bounded local memory queue drops oldest telemetry gracefully if offline for extended periods.