What Claude Code Is and What It Isn't

The failure it addresses

You SSH into a core switch. An interface is flapping. You check counters, see CRC errors climbing. You open a browser, search for the platform-specific transceiver diagnostic command. You find a Cisco forum post from 2019 that might apply to your IOS-XE version. You copy the command, paste it, get a syntax error because the post was for NX-OS. You search again. Twenty minutes later, you've typed three useful commands.

That cycle — intent, search, translate, validate, execute — is the actual bottleneck in network operations. Not the routing protocol convergence. Not the spanning tree calculation. The human lookup loop. Claude Code eliminates the lookup. You describe what you need. It produces the exact syntax for your platform and version. It runs the command if you let it. It reads the output and tells you what's broken.

What Claude Code actually is

Claude Code is an agentic CLI that runs in your terminal. It has tools: it reads files, writes files, executes bash commands, searches directories, and connects to external services via MCP. When you give it a task, it decides which tools to use, in what order, and chains them together. It's not autocomplete. It's not a chat interface with documentation knowledge. It's an agent that operates on your filesystem and your network — with your permission, under your supervision, constrained by rules you define.

Installation

# macOS, Linux, or WSL (recommended — native binary)
curl -fsSL https://claude.ai/install.sh | sh

# macOS via Homebrew
brew install --cask claude-code

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

# Verify
claude --version

# Start in your network automation project
cd ~/network-automation
claude

You need a Claude subscription: Pro ($20/month), Max ($100/month), or Max 20x ($200/month). Or an Anthropic API key for pay-per-use. The subscription is subsidized relative to raw API pricing. For daily operational use, Pro is sufficient to start; Max is worth it if you're running long troubleshooting sessions that consume context.

Your first networking prompt

Don't start ambitious. Start with something you'd normally Google:

You: What's the IOS-XE 17.x command to check transceiver
     optical power levels on TenGigabitEthernet1/0/1?

Claude: show interfaces TenGigabitEthernet1/0/1 transceiver detail

        This gives Tx/Rx power in dBm, temperature, voltage, and
        laser bias current with alarm thresholds.

        On older IOS-XE (pre-17.x), the command was:
        show hw-module subslot 1/0 transceiver 1 status

        The newer syntax is more reliable and includes DOM
        threshold alerts inline.

Three seconds. No browser. No version mismatch. No forum post from a different platform. Claude knows the version differences and tells you which command is better on modern code without being asked.

The three models

Claude Code gives you access to three model tiers. The difference matters for networking work:

Sonnet 4.6 — Your daily driver. Reads configs, generates templates, answers CLI syntax questions, interprets show command output. Handles 90% of operational tasks. This is the default.

Opus 4.6 — The heavy reasoner. Use for multi-vendor troubleshooting, architecture reviews, complex Terraform modules, or any task where Claude needs to hold large amounts of state (like correlating configs across twenty devices). Costs more tokens. Worth it for hard problems.

Haiku 4.5 — The fast lookup. Use for quick syntax checks, OID references, or simple config fragment generation. Fastest response time, cheapest per token. Not reliable for multi-step reasoning.

Switch models mid-session with /model:

/model sonnet     # Default — config analysis, template generation
/model opus       # Complex — multi-device correlation, architecture
/model haiku      # Quick — syntax lookup, simple transforms