Claude Code

Claude Code exports OpenTelemetry metrics and events itself, straight to Status200: no collector and no gateway in between. Turn it on with a few environment variables, or for a whole fleet with one managed settings file, and see cost, tokens, sessions, commits and accepted edits per developer.

How it works#

Claude Code sends:

• Cost and tokens: claude_code.cost.usage (USD) and claude_code.token.usage, by model, by main agent or subagent, and by the agent, skill, plugin or MCP server that made the call.
• Activity: sessions, commits, pull requests, lines of code added and removed, and active time.
• Edit decisions: claude_code.code_edit_tool.decision counts accepted and rejected edits, so the acceptance rate is accept / (accept + reject), by language and by tool.
• Events: prompts, tool results, API requests and errors, permission-mode changes and MCP server connections, as OTLP log records.

Every metric data point and event carries user.email when the developer is signed in with a Claude account, and Status200 reads it as the employee on the metrics, where the cost and tokens are.

Before you begin#

1. Create a telemetry ingestion key under Project Settings → Telemetry & APM → Ingestion Keys.
2. Your OTLP endpoint is https://status200.ru/otlp, or https://YOUR-ONEUPTIME-HOST/otlp if you host Status200 yourself.
3. Use one key for the whole fleet. It ends up on every developer machine, so treat it as a shared credential that can only send data, and rotate it like any other.

Set up one machine#

1. Export the variables#

   Add these to your shell profile (.zshrc or .bashrc), or export them in the terminal you start Claude Code from:

   Bash

   export CLAUDE_CODE_ENABLE_TELEMETRY=1
   export OTEL_METRICS_EXPORTER=otlp
   export OTEL_LOGS_EXPORTER=otlp
   export OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf
   export OTEL_EXPORTER_OTLP_ENDPOINT="https://status200.ru/otlp"
   export OTEL_EXPORTER_OTLP_HEADERS="x-status200-token=YOUR_INGESTION_TOKEN"

   Variable	What it does
   CLAUDE_CODE_ENABLE_TELEMETRY	The main switch: nothing is exported without it.
   OTEL_METRICS_EXPORTER, OTEL_LOGS_EXPORTER	otlp sends to Status200. otlp,console also prints what is sent, handy while you set up.
   OTEL_EXPORTER_OTLP_PROTOCOL	http/protobuf; http/json and grpc work too.
   OTEL_EXPORTER_OTLP_ENDPOINT	The base URL. Claude Code adds /v1/metrics and /v1/logs itself.
   OTEL_EXPORTER_OTLP_HEADERS	Comma-separated key=value pairs; here, your ingestion key.

2. Restart Claude Code#

   The variables are read when Claude Code starts. Metrics are sent every 60 seconds and events every 5 seconds, so use Claude Code for a minute before you look.

3. Check that data arrives#

   Open Metrics and look for claude_code.token.usage, then open Products → AI / LLM and go to Usage: your cost shows under your email, marked from GenAI metrics.

To send metrics and events to different places, set OTEL_EXPORTER_OTLP_METRICS_ENDPOINT and OTEL_EXPORTER_OTLP_LOGS_ENDPOINT instead, with the full path (https://status200.ru/otlp/v1/metrics, https://status200.ru/otlp/v1/logs).

Roll out to the fleet#

For an organization, ship the same variables in a managed settings file. It has the shape of settings.json, with the variables in an env block:

{
  "env": {
    "CLAUDE_CODE_ENABLE_TELEMETRY": "1",
    "OTEL_METRICS_EXPORTER": "otlp",
    "OTEL_LOGS_EXPORTER": "otlp",
    "OTEL_EXPORTER_OTLP_PROTOCOL": "http/protobuf",
    "OTEL_EXPORTER_OTLP_ENDPOINT": "https://status200.ru/otlp",
    "OTEL_EXPORTER_OTLP_HEADERS": "x-status200-token=YOUR_INGESTION_TOKEN",
    "OTEL_RESOURCE_ATTRIBUTES": "team.id=platform,cost_center=eng-123"
  }
}

1. Put the file where Claude Code reads it#

   Platform	Path
   macOS	/Library/Application Support/ClaudeCode/managed-settings.json
   Linux and WSL	/etc/claude-code/managed-settings.json
   Windows	C:\Program Files\ClaudeCode\managed-settings.json

   Claude Code no longer reads the old Windows path C:\ProgramData\ClaudeCode\managed-settings.json. If you already ship a managed settings file, add a drop-in instead, such as managed-settings.d/10-telemetry.json next to it: Claude Code merges managed-settings.json first, then every *.json file in managed-settings.d/ in alphabetical order.

2. Check one machine#

   Run /status in Claude Code. The Setting sources line should read Enterprise managed settings (file), or (drop-ins) or (file + drop-ins) for a drop-in. A different source in brackets means a higher-ranked source won, and yours was not used.

3. Watch the fleet arrive#

   On the Usage tab, each developer shows up as a row once they have used Claude Code with the new settings.

The same JSON can also come from server-managed settings in the claude.ai admin console (fetched at startup and every hour), a macOS configuration profile in the com.anthropic.claudecode domain, or a Settings value under HKLM\SOFTWARE\Policies\ClaudeCode on Windows.

Attribute usage to people and teams#

The employee is user.email. Claude Code sends it when the developer is signed in with a Claude account, and there is no variable to turn it off. Status200 reads it from every cost and token data point with no setup.

Other sign-ins send no email. With an API key, or with Claude on Amazon Bedrock, Google Cloud or Microsoft Foundry, there is no Claude account in the session, and Claude Code sends only user.id and session.id. Set the identity yourself, per machine, in the managed settings or a launch wrapper:

export OTEL_RESOURCE_ATTRIBUTES="user.email=jdoe@example.com,team.id=platform"

user.id is not a person. It is a random id Claude Code writes to ~/.claude.json on first run: one per install, and a new one if the file is deleted. Without an email, each install shows as its own row. One exception: a session signed in to a Claude apps gateway sends the signed-in identity (user.id is the identity provider's subject, user.email the email) and ignores user.* keys you set yourself.

Teams and cost centres come from OTEL_RESOURCE_ATTRIBUTES: Status200 reads team.id, team, cost_center and department. Values cannot contain spaces, and quotes do not escape them: write team=Platform_Engineering, or percent-encode, as in team=Platform%20Engineering. Claude Code copies these keys onto every data point (OTEL_METRICS_INCLUDE_RESOURCE_ATTRIBUTES, on by default), and Status200 reads them either way. A key you set never overrides one Claude Code sets itself.

What Status200 shows#

On Products → AI / LLM → Usage, claude_code.cost.usage feeds the Employee, Team and Model tables, and claude_code.token.usage the token totals. Two rules decide what you see:

• The metrics fill in for spans; they are never added to them. If the same project gets GenAI spans with cost from your own apps, Claude Code's spend does not show. Send Claude Code to a project of its own. See Spans come first, metrics fill in.
• A row from metrics carries cost only. Its calls and token cells show —, and the Provider and Application / Service tables stay empty: the metrics name no provider and belong to no Status200 service.

Everything else, such as acceptance rate, lines of code, commits and pull requests per developer, or spend by query_source, you chart on a dashboard from the metrics below, and alert on with a Metrics Monitor.

Metrics#

Every metric is a counter, and carries the standard attributes as well as its own.

The standard attributes are session.id, app.version, app.entrypoint, organization.id, user.account_uuid, user.account_id, user.id, user.email and terminal.type, plus your OTEL_RESOURCE_ATTRIBUTES. Some of them can be turned off: see Keep metric volume down.

• The cache token types are camelCase: cacheRead and cacheCreation. Status200's token totals count only input and output.
• On the cost and token metrics, the names of your own agents and MCP servers, and of third-party plugins and skills, read custom or third-party unless OTEL_LOG_TOOL_DETAILS=1 is set (Claude Code 2.1.273 or later).

Events#

Events arrive as OTLP log records, each with event.name, event.timestamp and event.sequence. The ones you will use most:

Claude Code sends many more, such as refusals, permission-mode changes and plugin, hook and skill events: Anthropic's monitoring reference lists them all. In cost_usd_micros, cost is a whole number of millionths of a USD, which adds up without rounding drift.

Keep metric volume down#

Claude Code puts session.id on every data point by default, and every session is a new series. Across a fleet that one attribute drives metric volume, and cost reporting doesn't need it:

export OTEL_METRICS_INCLUDE_SESSION_ID=false

user.email, user.id, organization.id and terminal.type have no switch. Events are the busier signal: each active developer flushes every 5 seconds (OTEL_LOGS_EXPORT_INTERVAL, in milliseconds). Raise it to send fewer, larger batches, or set OTEL_LOGS_EXPORTER=none if you only need cost and usage, which all live in the metrics.

Prompt and tool content are off by default#

Claude Code does not export prompts, answers, tool arguments or tool output unless you turn each one on:

We recommend leaving them off: the metrics and the redacted events already answer who spent what, on which model, with what acceptance rate. Content capture moves prompts, source code and tool output into a telemetry store with different access controls than your repositories. If you turn it on for an investigation, CLAUDE_CODE_OTEL_CONTENT_MAX_LENGTH (61440 by default) limits how much travels per record, and the traces' Scrub Rules and Drop Filters under Traces → Settings mask or drop it before it is stored.

Metrics arrive as deltas#

Claude Code sends delta metrics by default: each data point is the increase since the last export, not a running total. When you chart or alert on them, add the data points up over the window, and don't apply a rate or counter-reset function as you would to a cumulative counter: the numbers come out wrong without looking broken.

export OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE=cumulative

Pin cumulative only if every exporter in your organization uses it, and don't mix the two across a fleet.

Traces (beta)#

Claude Code also has a beta traces exporter, which draws each interaction as a span tree: claude_code.interaction, with claude_code.llm_request and claude_code.tool beneath it.

export CLAUDE_CODE_ENHANCED_TELEMETRY_BETA=1
export OTEL_TRACES_EXPORTER=otlp

In Agent SDK and non-interactive -p sessions, Claude Code picks up an incoming TRACEPARENT, so its spans join the caller's trace; interactive sessions ignore it. It sends the traceparent header to the model only when ANTHROPIC_BASE_URL is unset or the Anthropic API; behind your own proxy, set CLAUDE_CODE_PROPAGATE_TRACEPARENT=1.

Troubleshooting#

Next steps#