MCP Server Health Monitoring for Production Agents

Production AI agents using MCP servers have no built-in observability. When search quality degrades, there is no way to tell whether the MCP server is down, the upstream provider is rate-limited, or the agent is making poor tool selections. Teams discover problems only when users complain about bad answers. That is too late.

What to monitor

Four metrics per MCP tool: success rate, latency, result count, and error type. Set baselines for each. Google search typically returns in 1-2 seconds with 10+ results. Reddit search may be slower with fewer results. YouTube search has different characteristics still. Each tool needs its own baseline, not a single threshold applied globally.

The health check pattern

import requests, os, time, json

H = {'x-api-key': os.environ['SCAVIO_API_KEY']}
PLATFORMS = ['google', 'reddit', 'youtube', 'amazon', 'walmart']

def health_check():
    report = {'ts': time.strftime('%Y-%m-%dT%H:%M:%SZ')}
    for p in PLATFORMS:
        start = time.time()
        try:
            resp = requests.post('https://api.scavio.dev/api/v1/search',
                headers=H, json={'platform': p, 'query': 'test'}, timeout=15)
            report[p] = {
                'ok': True,
                'latency': round(time.time() - start, 2),
                'results': len(resp.json().get('organic', [])),
                'status': resp.status_code
            }
        except Exception as e:
            report[p] = {'ok': False, 'error': str(e)}
    return report

Routing informed by health data

Health data is not just for alerting. Feed it back into routing decisions. If YouTube search latency spikes above 5 seconds, the agent can temporarily prefer Google search for video queries. If Amazon search returns zero results for a period, route product queries to Walmart. Health-aware routing turns monitoring from passive observability into active resilience.

Run every 5 minutes

A 5-minute health check cadence catches most issues before users notice. Log to JSONL for historical analysis. Alert to Slack for immediate visibility. The 5 API calls per check cost $0.025. That is $7.20 per month for comprehensive MCP health monitoring. Cheaper than a single customer escalation caused by undetected degradation.