A Grafana dashboard for monitoring Claude Cowork for Team usage via OpenTelemetry.
Cowork sends telemetry events to Grafana Cloud (Loki) through its built-in OTLP integration. This dashboard visualizes that data using LogQL queries.
- Overview — Total API requests, cost (USD), prompt count, error count
- User Activity — Per-user API request and cost trends (grouped by
user_email) - AI Utilization — Sessions per user, prompts per user, avg prompts per session, Skill/MCP tool usage ranking, sub-agent usage
- Best Practices — Session split appropriateness (max vs avg prompts), personal skill growth, provided Skill adoption rate
- Cost & Tokens — Cost breakdown by model, token usage (input/output/cache), speed mode distribution, API response duration (avg/max)
- Tool Usage — Top tools ranking, tool accept/reject decisions
- Harness Quality — MCP tool error rate, tool rejection rate, provided Skill utilization
- Errors — Error count timeline and recent error log viewer
- Claude Cowork for Team with OpenTelemetry enabled (Admin settings > Cowork > Monitoring)
- Grafana Cloud account with Loki data source
In the Cowork admin panel, set:
| Setting | Value |
|---|---|
| OTLP Endpoint | Your OTLP endpoint shown in Grafana Cloud Portal > Stack > Details > OpenTelemetry |
| OTLP Protocol | http/protobuf |
| OTLP Headers | Authorization=Basic <base64(INSTANCE_ID:API_TOKEN)> |
- Open Grafana UI
- Go to Dashboards > New > Import
- Upload
dashboard.jsonor paste its contents - Select your Loki data source when prompted
The dashboard uses a template variable ${logsDs} for the Loki data source. You will be prompted to select it on import. Cowork events are stored with service_name=cowork.
- Cowork exports events/logs only (not metrics). All queries are LogQL-based against Loki.
- Event fields (cost, tokens, user info, etc.) are stored as Loki structured metadata, not as JSON in the log body. Queries reference metadata fields directly without
| json. - The Claude Code Stats Grafana plugin requires Prometheus metrics and does not work with Cowork log data.
MIT