Claude Code OpenTelemetry: Track Cost per Developer and Repo

Claude Code can export its own usage over OpenTelemetry, including an estimated dollar cost for every model request. Turn on telemetry, enable the logs exporter, set a protocol, and point it at a receiver. Then each developer’s spend arrives as claude_code.api_request events carrying cost_usd, the model, and token counts. Add OTEL_METRICS_INCLUDE_REPOSITORY=true and each event also names the repository. That’s enough to total Claude Code spend per developer and per repo with nothing extra installed.

This guide covers the settings, the fields that matter for cost, the gotchas, and where to send the data.

The settings

In ~/.claude/settings.json (or as environment variables):

{
  "env": {
    "CLAUDE_CODE_ENABLE_TELEMETRY": "1",
    "OTEL_LOGS_EXPORTER": "otlp",
    "OTEL_EXPORTER_OTLP_PROTOCOL": "http/protobuf",
    "OTEL_EXPORTER_OTLP_ENDPOINT": "https://your-collector.example.com",
    "OTEL_EXPORTER_OTLP_HEADERS": "Authorization=Bearer your-token",
    "OTEL_METRICS_INCLUDE_REPOSITORY": "true"
  }
}

Setting	What it does
CLAUDE_CODE_ENABLE_TELEMETRY	Turns telemetry on. Off by default.
OTEL_LOGS_EXPORTER	Exports events, including one api_request per model call.
OTEL_EXPORTER_OTLP_PROTOCOL	grpc, http/json, or http/protobuf. There is no default.
OTEL_EXPORTER_OTLP_ENDPOINT	Base URL. HTTP exporters append /v1/logs and /v1/metrics.
OTEL_EXPORTER_OTLP_HEADERS	Auth for the receiver. Use one token per developer if the receiver attributes by token.
OTEL_METRICS_INCLUDE_REPOSITORY	Adds the repository from the origin remote. Off by default; Claude Code 2.1.269 or later.

To roll this out to a whole team, put the shared settings in a managed settings file instead of asking each developer to edit theirs. If your receiver attributes spend by token, keep the header out of the managed file, since one shared token would merge everyone into one person; each developer sets their own. See Anthropic’s monitoring docs.

The fields that matter for cost

Each claude_code.api_request event carries:

• cost_usd: estimated cost of the request in USD, and cost_usd_micros as an integer
• model
• input_tokens, output_tokens, cache_read_tokens, cache_creation_tokens (input excludes cached tokens)
• session.id and prompt.id
• user.email when the developer is signed in to a Claude account, and user.id, a random per-install id
• vcs.owner.name and vcs.repository.name, when the repository setting is on

Summing cost_usd grouped by user.email (or by token) gives cost per developer. Don’t group by user.id: it’s an anonymous id that changes if ~/.claude.json is deleted. Grouping by vcs.repository.name gives cost per repo. Grouping by session.id gives cost per session.

The claude_code.cost.usage metric carries the same spend as a counter on a 60-second interval. It’s fine for a running total on a chart. The per-request event is better for attribution because it ties each dollar to a session, a prompt, and a repository.

Gotchas

• No default protocol. Without OTEL_EXPORTER_OTLP_PROTOCOL, the exporter doesn’t send. Check your receiver supports the one you pick: many hosted HTTP endpoints don’t serve grpc.
• Restart required. Settings are read at startup.
• Noise. One prompt can produce dozens of events: hooks registered, MCP connections, tool results. If you store everything, keep api_request, api_error, and user_prompt, and drop the rest.
• Repo attributes are off by default, and they’re omitted when the session has no origin remote.
• No branch or pull request. Telemetry names the repo but not the branch, so cost per PR needs another source.
• Estimates. cost_usd is an estimate at list prices. It won’t match an invoice that has seat plans or discounts.
• Claude Code only. Cursor’s OpenTelemetry export is limited to Enterprise. Codex and Gemini CLI need their own setup, so a mixed-tool team won’t get one view from this alone.

Where to send it

• Your own stack. An OpenTelemetry Collector into Grafana, Honeycomb, Datadog, or any OTLP backend. You build the queries and dashboards yourself.
• PromptConduit. Point the endpoint at https://api.promptconduit.dev/v1/otlp with your API key, and Claude Code spend counts on the team page. It’s broken down by member, model, repository, client, and day, with budgets, email and Slack alerts, and a savings report. Setup is in the Claude Code telemetry docs. It accepts http/protobuf and http/json, gzip or not.

For branch and pull request attribution, and for Cursor, Codex, and Gemini CLI in the same report, add the CLI. Where both send the same session, the CLI’s figures are used and the telemetry cost is left out, so nothing is counted twice. See how to attribute AI coding spend to teams and repos.