Back
December 20, 2024 ai

Building AI-Powered Network Diagnostics

What if an AI could diagnose network issues faster than your best network engineer? That’s the question we set out to answer.

The Problem Space

Our support team handles hundreds of network-related tickets per week. “WiFi is slow.” “POS can’t connect.” “Kitchen display is offline.” Each one requires someone to:

  1. Log into the ticketing system
  2. Look up the location’s network devices
  3. SSH into routers and access points
  4. Run various diagnostic commands
  5. Interpret the results
  6. Formulate a response

Average time: 15-20 minutes per ticket. Multiply that by 400 tickets a week, and you’re looking at 100+ hours of skilled engineer time just on diagnostics.

Enter MCP

Model Context Protocol (MCP) is Anthropic’s standard for connecting AI models to external tools. Instead of fine-tuning a model on network diagnostics, we give it access to the same tools our engineers use.

The architecture is simple:

User describes problem
        ↓
Claude analyzes and decides what to check
        ↓
MCP Server executes network API calls
        ↓
Results returned to Claude
        ↓
Claude interprets and responds

Building the MCP Server

Our MCP server exposes tools for:

  • Device lookup - Find all network devices at a location
  • Ping tests - Check connectivity between devices
  • Throughput tests - Measure actual bandwidth
  • Client enumeration - See what’s connected to the network
  • Event logs - Check recent network events

Each tool maps to an API call to either Meraki (our wireless vendor) or Pronto (our router management platform).

The Results

Early testing shows the AI can correctly diagnose common issues about 80% of the time. More importantly, it does it in under 60 seconds instead of 15 minutes.

The types of issues it handles well:

  • ISP outages - Detects when the WAN link is down
  • WiFi interference - Identifies channel congestion
  • Client density - Spots when too many devices are on one AP
  • DHCP exhaustion - Recognizes when IP pools are full

What’s Next

We’re working on:

  1. Automated remediation - For simple fixes, let the AI take action
  2. Symptom pattern matching - Learn from past tickets to get smarter
  3. Proactive alerting - Detect issues before they become tickets

The goal isn’t to replace network engineers. It’s to handle the routine 80% so engineers can focus on the interesting 20%.