---
title: Send your agent's own telemetry to sigiro
description: Claude Code and Codex both emit OpenTelemetry about their own sessions. Point them at sigiro to read token counts, tool calls and session latency back with SQL.
sidebar:
  order: 3
---

This guide shows you how to record what your agent does, rather than what your
service does. Claude Code and Codex both emit OpenTelemetry about their own
sessions. Point that telemetry at sigiro, and the sessions become rows you can
query.

This is a different task from
[Instrument your code with an agent](/docs/how-to/install-with-ai), which points
_your service_ at sigiro. Do both if you want both.

## Before you start

You need a sigiro server. Start one with `sigiro serve`, or with the Docker
command in the [quickstart](/docs/tutorials/quickstart#1-run-it).

Both agents export over OTLP gRPC on port `4317` in the settings below. A
self-hosted server needs no key.

## Claude Code

Claude Code reads its telemetry settings from the environment. Set these
variables in the environment that starts Claude Code. Use a shell profile, an
`.envrc` file, or a wrapper script:

```bash
export CLAUDE_CODE_ENABLE_TELEMETRY=1
export CLAUDE_CODE_ENHANCED_TELEMETRY_BETA=1
export OTEL_METRICS_EXPORTER=otlp
export OTEL_LOGS_EXPORTER=otlp
export OTEL_TRACES_EXPORTER=otlp
export OTEL_EXPORTER_OTLP_PROTOCOL=grpc
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
```

For the hosted service, add
`OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer ${SIGIRO_API_KEY}"`. Keep the
key out of every committed file.

## Codex

Codex reads its telemetry settings from its own configuration file, not from the
environment. Edit `~/.codex/config.toml`:

```toml
[otel]
environment = "dev"
log_user_prompt = false
exporter = "otlp-grpc"

[otel.exporter."otlp-grpc"]
endpoint = "http://localhost:4317"
```

For the hosted service, add
`headers = { authorization = "Bearer ${SIGIRO_API_KEY}" }`. Set `SIGIRO_API_KEY`
in the environment that starts Codex.

## One limit applies to both

These settings cover the agent process only. They do not cover the commands that
the agent runs. An application that the agent starts through its shell tool does
not always inherit these variables, so an instrumented service still needs its
own exporter settings. Use
[Instrument your code with an agent](/docs/how-to/install-with-ai) for that.

## Confirm that it works

Start the agent, run one prompt, then ask sigiro which services it received:

```bash
curl http://localhost:9999/v1/services
```

The agent reports itself as a service. After it appears, query the sessions:

```bash
curl -X POST http://localhost:9999/v1/query \
  --data "SELECT service_name, span_name, count(*) AS n
          FROM sigiro_spans
          WHERE timestamp > now() - INTERVAL '1 hour'
          GROUP BY 1, 2 ORDER BY n DESC"
```

An agent also reports counters and gauges, and those land in the
`sigiro_metrics_*` tables rather than in `sigiro_spans`. [About the
tables](/docs/explanation/tables) says which family holds what, and which
machine each family describes.

## Next

- [Instrument your code with an agent](/docs/how-to/install-with-ai) — point your
  service at sigiro as well
- [CLI reference](/docs/reference/cli) — every subcommand and flag