OpenTelemetry

ClawMetry speaks OpenTelemetry in both directions, using the GenAI semantic conventions, so your agent traces are never locked into one tool.

Export every session — LLM calls, tools, sub-agents, tokens, cost — as OTLP/HTTP GenAI spans to any collector (Datadog, Grafana, Honeycomb, or your own OTel Collector):

clawmetry --otel-export http://localhost:4318/v1/traces
# equivalently:
CLAWMETRY_OTEL_EXPORT_ENDPOINT=http://localhost:4318/v1/traces clawmetry

Auth headers and poll interval are optional env vars:

CLAWMETRY_OTEL_EXPORT_HEADERS='{"X-API-Key":"…"}'   # extra HTTP headers
CLAWMETRY_OTEL_EXPORT_INTERVAL=60                    # seconds (default 60)

Ingest — the built-in OTLP receiver accepts traces, logs, and metrics from anything else at /v1/traces, /v1/logs, and /v1/metrics. Point any OpenTelemetry-instrumented app at it:

OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:8900 OTEL_EXPORTER_OTLP_PROTOCOL=http/json your-app

OTLP/JSON traces, logs and metrics work on a plain pip install clawmetry, no extras. Protobuf ingest needs pip install clawmetry[otel]. An app that sets its own service.name shows up as its own agent in the runtime switcher, with its cost and tokens.

Switch a runtime's own exporter on, one command. Several runtimes ship an OpenTelemetry exporter of their own (Claude Code, Codex, Gemini CLI, Cursor, Copilot, OpenCode and more), each off by default and each configured differently. clawmetry instrument <runtime> writes that runtime's settings so its exporter reports to the local receiver, and --uninstall removes exactly what it wrote. The mechanics are shared: it merges into the existing settings, never overwrites a value it did not write, keeps a per-file record of its keys, refuses when managed policy pins the destination, and keeps prompt and tool content OFF unless you pass --content. Raw request and response bodies are never enabled.

Which runtimes are available depends on which exporter profiles are registered. A profile is what tells the vendor-neutral receiver a runtime's metric names, event names, span attribute spellings, session-id form and settings file. Free runtimes' profiles ship in this repo; paid runtimes' profiles ship in the clawmetry-pro wheel alongside their transcript adapters, and register through the same plugin entry point. clawmetry instrument --help lists what is registered on this install. With no profile registered, an emitter is treated as any other OpenTelemetry app: spans, ledger rows and the generic tiles, nothing runtime-specific.

Claude Code, for example: its profile (in clawmetry-pro) writes the env block below into ~/.claude/settings.json (--project for the repo-local file). Claude Code has no default protocol, so the protocol key is required; http/json works on a vanilla install.

{
  "env": {
    "CLAUDE_CODE_ENABLE_TELEMETRY": "1",
    "OTEL_LOGS_EXPORTER": "otlp",
    "OTEL_METRICS_EXPORTER": "otlp",
    "OTEL_TRACES_EXPORTER": "otlp",
    "CLAUDE_CODE_ENHANCED_TELEMETRY_BETA": "1",
    "OTEL_EXPORTER_OTLP_PROTOCOL": "http/json",
    "OTEL_EXPORTER_OTLP_ENDPOINT": "http://127.0.0.1:4318",
    "OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE": "delta"
  }
}

Restart the runtime afterwards: a running session keeps the configuration it started with. clawmetry status and GET /api/otel-status (runtimes) report, per profiled runtime, whether the block is in place and when its last batch arrived.

What a profiled runtime's exporter adds over its transcript: permission decisions and their source, permission-mode changes, API refusals and errors, MCP server connection health, time spent waiting on you, lines of code, commits and pull requests, and per-skill, per-agent and per-MCP-server attribution on every request, all joined to the same session the transcript adapter builds.

You get the zero-config, local-first ClawMetry dashboard and your data in whatever backend your team already runs — no lock-in, no second agent to install.

The scheduled push exporter (Pro) has its own page: OTEL_PUSH_EXPORTER.md.