Cloud Monitoring Metric Selection Agent Skill

Metric Selection (Service Query & Local Keyword Filtering)

Use this skill to identify the most relevant Google Cloud Monitoring metric

descriptors. It queries all metric descriptors for a target service from the API

and filters them locally inside the agent's context using keyword matching.

CRITICAL RULES

• Always Query Live APIs: You MUST always retrieve the most up-to-date

metric descriptors dynamically by calling the list_metric_descriptors MCP

tool.

Workflow

Step 1: Verify & Auto-Configure MCP

• Check if any tool matching list_metric_descriptors (e.g.

google-cloud-monitoring:list_metric_descriptors,

mcp_google-cloud-monitoring_list_metric_descriptors, or a similar pattern)

is available in your active toolset.

• Verify via Unique URL: To ensure you are calling the correct Google

Cloud Monitoring tool, confirm that the underlying MCP server configuration

points to: https://monitoring.googleapis.com/mcp.

• If the tool is missing:

• Locate the MCP configuration file for the user's environment. Check

common paths:

• ~/.gemini/config/mcp_config.json
• ~/.codeium/windsurf/mcp_config.json
• cline_mcp_settings.json
• claude_desktop_config.json
• Directly update/merge the configuration file with the following server

configuration. CRITICAL: Merge the JSON object to preserve any

existing MCP servers in mcpServers. Do not overwrite the file.

        "google-cloud-monitoring": {
          "url": "https://monitoring.googleapis.com/mcp",
          "authProviderType": "google_credentials",
          "enabledTools": [
            "list_metric_descriptors"
          ]
        }

• Print a clear message notifying the user that the

google-cloud-monitoring MCP server has been configured, and request

them to restart or start a new chat session to refresh tools. Stop

calling further tools and end the turn.

Step 2: Analyze Request & Extract Keywords

• Identify the target GCP service prefix (e.g. compute, spanner,

bigquery, storage) and the project ID from the resource URI.

• Extract target metric concepts from the user's prompt (e.g., "CPU",

"memory", "bytes scanned", "latency", "connections").

• Map these concepts to standard Google Cloud Monitoring metric substrings

(e.g., cpu, mem, scanned_bytes, latenc, connections).

*Example Query Analysis:*

• User Prompt: "Check Cloud Storage bucket write throughput and request

count"

• Resource URI:

//storage.googleapis.com/projects/my-project/buckets/my-bucket

• Service Prefix: storage (mapped to storage.googleapis.com)
• Metric Keywords: write, throughput, request, count
• Mapped Substrings: write, throughput, request_count, count

Step 3: Query Metric Descriptors via list_metric_descriptors Tool

Query all metric descriptors for each identified service prefix using the

list_metric_descriptors MCP tool (using pageSize: 200). Because Google Cloud

Monitoring filters do not allow combining multiple metric.type restrictions

with OR, you must **initiate a separate query for each identified service

prefix** (either sequentially or in parallel).

If any response includes a nextPageToken, you MUST make consecutive follow-up

calls passing pageToken until all remaining descriptors for that prefix are

retrieved before filtering.

*Filter Pattern Construction:* Map the target service domain to its appropriate

prefix style:

• Standard Google Cloud Services:

starts_with("<service_prefix>.googleapis.com/") (e.g.,

bigquery.googleapis.com/, redis.googleapis.com/).

• Ops Agent (Guest OS): starts_with("agent.googleapis.com/") (for guest

OS memory/disk metrics).

• Kubernetes / GKE Native: starts_with("kubernetes.io/")
• Istio Service Mesh: starts_with("istio.io/")
• Knative Serving / Autoscaler: starts_with("knative.dev/")
• Custom / External Metrics: Use starts_with("custom.googleapis.com/")

or starts_with("external.googleapis.com/").

*Example Tool Call Payload:* If both Spanner and Compute Engine are targeted in

the request, execute these two tool calls:

• Spanner query:

{
  "name": "projects/my-project-id",
  "filter": "metric.type = starts_with(\"spanner.googleapis.com/\")",
  "pageSize": 200
}

• Compute Engine query:

{
  "name": "projects/my-project-id",
  "filter": "metric.type = starts_with(\"compute.googleapis.com/\")",
  "pageSize": 200
}

Call the list_metric_descriptors tool with these payloads.

Step 4: Local Filtering & Fallback Protocol

Aggregate all descriptors returned from Step 3, and filter them locally inside

your LLM context:

• Keyword Filtering: Filter the list by matching your target metric

keywords (e.g. "cpu", "latency") against the type, displayName, and

description fields of the descriptors.

• Resource Alignment: Check if the metric contains labels matching the

target resource granularity (e.g., checking for a database label if

targeting a database resource). Do not attempt to dynamically match resource

type strings directly, as Google Cloud Monitoring resource mappings (like

Spanner databases mapping to spanner_instance) can be counter-intuitive.

Troubleshooting & API Fallbacks

If any tool call fails, times out, or returns empty results, use these

strategies:

• Case A: API Syntax Error: Examine the error message, correct the filter

syntax, and retry.

• Case B: Timeout / Rate Limits: Retry the call once with a smaller page

size (e.g., pageSize: 20).

• Case C: Unrecoverable Failure / Empty List:
• Verify if the target service is enabled in the project.
• Search Google Cloud public documentation to verify standard metrics for

the service.

• Notify the user of the failure and ask for clarification.

Step 5: Output Selected Metrics

For each service domain, return only the 5-15 key metrics directly relevant to

the user's intent.

You MUST report the selected metrics in clean Markdown tables, grouped by

service (i.e., one table per service prefix). The table MUST include the

following columns: "Metric Type", "Display Name", "Description", "Metric Kind",

"Value Type", "Unit", and "Monitored Resource Types". Map the fields from the

Google Cloud Monitoring list_metric_descriptors tool call response objects

directly to the table columns:

• Metric Type: Map to the type field (e.g.,

spanner.googleapis.com/instance/cpu/utilization).

• Display Name: Map to the displayName field.
• Description: Map to the description field.
• Metric Kind: Map to the metricKind field (e.g., GAUGE, DELTA,

CUMULATIVE).

• Value Type: Map to the valueType field (e.g., INT64, DOUBLE,

DISTRIBUTION, BOOL).

• Unit: Map to the unit field (e.g., 1, By, s, ms).
• Monitored Resource Types: Map to the monitoredResourceTypes list field

(e.g., ["spanner_instance"]).

*Example Output Table:*

Metric Type | Display Name | Description | Metric Kind | Value Type | Unit | Monitored Resource Types

:------------------------------------------------ | :----------------------- | :------------------------------------------ | :---------- | :--------- | :--- | :-----------------------

spanner.googleapis.com/instance/cpu/utilization | Instance CPU Utilization | Fraction of allocated CPU currently in use. | GAUGE | DOUBLE | 1 | ["spanner_instance"]

Reference Documentation & Links

• Google Cloud Monitoring Metric List:

GCP Metrics Documentation

• MetricDescriptor MCP Tool Reference:

MCP Tools Reference: monitoring.googleapis.com

• Monitoring Filter Syntax Guide:

Monitoring Filters