Telemetry export for AgentLens

AgentLens reads the local session files of each client directly. OpenTelemetry (OTel) is an optional second source: it adds what the local files do not hold, such as the cost Claude Code computed per request and the skill each request is attributed to.

What AgentLens imports

agentlens import-otel FILE reports what an export holds, and --otel FILE (repeatable, works with every command) adds it to an analysis.

Client	Imported	Notes
Claude Code	Log events claude_code.user_prompt, claude_code.api_request, claude_code.skill_activated, claude_code.tool_result	Measured tokens and cost per request; skill.name of a request gives measured per-skill attribution (the attributed column of agentlens skills).
Copilot CLI	Spans of the file exporter: chat (one model request with its tokens), execute_tool, invoke_agent (user turns)	Measured tokens per request, premium requests (github.copilot.cost) and subagent nesting. See Copilot CLI.
VS Code Copilot	The same spans as Copilot CLI	VS Code uses the same agent host, so the file has the same schema. See VS Code.

Not imported: metrics (claude_code.token.usage, the gen_ai.client.* metrics of Copilot and others) are aggregates without a request or timestamp. Prompts, responses and tool arguments are never read.

A session that is in the local Claude Code or Copilot CLI data and in an export (same session.id, for Copilot CLI gen_ai.conversation.id) is counted once. The export's requests (with the cost and skill attribution Claude Code measured) replace the local usage inside the time range the export covers; local requests outside that range stay, so a partial export does not undercount. Tool calls, skills, compactions, context items, project and branch come from the local copy.

Claude Code

Claude Code has no file exporter. It sends OTLP to a collector, and the collector writes the file.

1. Enable telemetry for the logs signal:

   export CLAUDE_CODE_ENABLE_TELEMETRY=1
   export OTEL_LOGS_EXPORTER=otlp
   export OTEL_EXPORTER_OTLP_PROTOCOL=grpc
   export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
   export OTEL_LOGS_EXPORT_INTERVAL=5000

2. Optional: export OTEL_LOG_TOOL_DETAILS=1 makes skill names appear verbatim. Without it, user-defined and third-party plugin skills are reported as custom_skill or third-party, and AgentLens attributes usage to those placeholder names.

3. Run an OpenTelemetry Collector (the contrib distribution contains the file exporter) with this configuration:

   receivers:
     otlp:
       protocols:
         grpc:
           endpoint: 127.0.0.1:4317
   exporters:
     file/claude:
       path: ./claude-code-logs.jsonl
   service:
     pipelines:
       logs:
         receivers: [otlp]
         exporters: [file/claude]

   The file exporter writes one OTLP JSON export per line, which is what AgentLens reads.

4. Analyze:

   agentlens import-otel claude-code-logs.jsonl
   agentlens skills --otel claude-code-logs.jsonl --since 30d

Privacy

• Prompts are redacted by Claude Code unless OTEL_LOG_USER_PROMPTS=1 is set. AgentLens never stores prompt text, tool parameters or commands, even if the export contains them; it keeps counts, sizes, names and timings only.
• With OTEL_LOG_TOOL_DETAILS=1 the export contains commands and tool input. Keep such files private.
• The collector configuration above listens on localhost only.

Sources

The attribute names come from the Claude Code monitoring documentation (https://code.claude.com/docs/en/monitoring-usage).

Copilot CLI

Copilot CLI writes JSON lines itself, no collector is needed:

mkdir -p /tmp/copilot-otel
COPILOT_OTEL_FILE_EXPORTER_PATH=/tmp/copilot-otel/trace.jsonl copilot
agentlens import-otel /tmp/copilot-otel/trace.jsonl
agentlens --otel /tmp/copilot-otel/trace.jsonl --source copilot-cli compare

Setting the path switches OpenTelemetry on and selects the file exporter. copilot help monitoring lists all variables. Message content (OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT) is off by default and not needed.

Each line is a span or a metric. AgentLens reads the spans:

Span	Becomes
chat <model>	A request with measured tokens. gen_ai.usage.input_tokens includes the cache read and write tokens, so the uncached input is the difference. github.copilot.cost is the number of premium requests.
execute_tool <name>	A tool call. A span status of 0 means unset, so the call is unknown, never a success; status 2 is an error.
invoke_agent	The github.copilot.user.message events become user turns. An invoke_agent nested inside another one marks its requests as subagent requests (inferred from the nesting; not yet seen in a real export).

Recent Copilot CLI versions also keep these numbers in their local session-store.db, and AgentLens reads them from there. The export then changes nothing: the request of the export replaces the local one of the same moment (the two differ by a few milliseconds, so the match allows 250 ms). It adds requests for sessions the store holds no rows for, and sessions that are only in the export.

VS Code

The Copilot agent in VS Code writes the same span schema as Copilot CLI. Switch the file exporter on in the settings (or start VS Code with the COPILOT_OTEL_* variables from a terminal), reload the window and chat:

{
  "github.copilot.chat.otel.enabled": true,
  "github.copilot.chat.otel.exporterType": "file",
  "github.copilot.chat.otel.outfile": "/tmp/copilot-otel/vscode.jsonl"
}

agentlens import-otel /tmp/copilot-otel/vscode.jsonl

The exporter buffers: the file stays empty until a few spans are flushed, so close the window if nothing shows up. Next to the chat, execute_tool and invoke_agent spans the file holds many host spans (github.copilot.session.timing.*, vscode.agent_host.*) and lines without a type. They carry no tokens and are reported as "other spans" and "unknown line", which is expected. The sessions are copilot-cli sessions, the same as the CLI's.