The answer
The three OpenTelemetry resource attributes that make several servers readable are service.name for the server, deployment.environment.name for production or staging, and service.version for the release. Without them, two servers that both expose a search tool merge into one misleading line.
| Attribute | Separates | Example |
|---|---|---|
| service.name | One server from another. | crm-mcp, docs-mcp |
| deployment.environment.name | Production from test traffic. | production, staging |
| service.version | One release from the next. | 2.4.0 |
Start each review per server: users, returning users, clients, tool usage and failures. Then compare servers on the same measures and the same window. To recognize one person across several servers, every server must send the same stable, permitted user identifier.
See MCP server analytics for the per-server view and the MCP monitoring guide for the full instrumentation checklist.
What should a multi-server MCP view show?
- One row per server with calls, error rate, p95 latency, users and returning users over the same window.
- The AI client mix per server, so a drop in one client is not hidden by growth in another.
- Top tools per server, plus tools that were never called, which often point to unclear descriptions.
- Release markers from service.version, so a change in one server is easy to line up with its deploy.
- One alert policy and one owner per server, instead of a single global error threshold.
- A shared, permitted user identifier across servers, so one person's journey across them stays visible.
Published