ClaudeCode_VMinfra_EN

What Can Claude Code Do When Given Admin Access to a Virtualization Platform? JANOG57.5 Interim Meeting · 2026-06-19 Ken Sasaki

Quick Poll for the Audience 1 Are you using generative AI? 2 Are you using Claude Code or Codex? 3 Have you used Claude Code for server / network operations? 4 Are you already using it in production?

Today's Topic Vibe Coding on a Server Virtualization Platform Building servers using natural language only What is Vibe Coding? A style of development where you don't think about the technical details. You just tell the AI in natural language — "make me an app that feels like this" or "fix this bug" — and let it handle everything. No code. Just instructions.

Test Environment: Virtualization Platform (ThinkPad X395) Hardware Division of Roles CPU AMD Ryzen 5 PRO 3500U RAM 16 GB 512 GB SSD Software • Ubuntu 24.04.04 Server • KVM / LXD / Docker ready • Storage managed with LVM Human's job • Set up SSH access for Claude Code • Give instructions in Japanese • Keep saying "OK" • Verify results, complain if broken Claude Code's job Everything else

Systems Built by Claude Code Prometheus + Grafana + Blackbox + Loki NetBox (+ Shumoku) Monitoring, visualization & log collection Infrastructure management & cable visualization containerlab OpenStack (DevStack) Network emulation environment Open-source IaaS platform

01 Prometheus + Grafana + Blackbox + Loki Modern monitoring stack

Prometheus + Grafana + Blackbox + Loki Prometheus Grafana Blackbox Exporter Loki Metrics collection (numerical data) Visualization & dashboards External monitoring (Ping, HTTP checks, etc.) Log collection & search Config files use YAML / INI format — tedious to write and debug manually…

Instructions Given to Claude Code (Monitoring) H I want to monitor my servers — what do you recommend? H OK, set it up with Prometheus then H Also add ping and HTTP monitoring H I also want hardware monitoring H And collect logs too H Create a dashboard for all of it

Grafana: Node Exporter Dashboard

Grafana + Loki: Log Collection Dashboard

Grafana: Hardware Health Dashboard (CPU temp, fans, battery)

Impressions: Monitoring Stack What worked well • • Did everything without us writing a single config line! Faster than a human When it got stuck, it mostly noticed and fixed itself Needed human nudges • Dashboard wasn't created initially • Log collection wasn't working at first • Hardware metrics weren't being collected → Fixed itself after being pointed out

02 NetBox (+ Shumoku) Infrastructure management & cable visualization

What is NetBox + Shumoku? NetBox A DCIM/IPAM tool for centrally managing infrastructure info — IP addresses, racks, devices, cable connections, etc. Easy to install, but entering and updating data is very tedious. Shumoku A tool that visualizes cabling based on NetBox data. Requires not just connection info, but proper 'cable' data to be entered — or it won't render.

Instructions Given to Claude Code (NetBox) H Set up a NetBox environment H I want to demo it — grab device templates from GitHub and install them H Also populate it with realistic demo data H There's a cable visualization tool called Shumoku — install that too and make it work

NetBox: Rack Diagram (Tokyo DC1 / Osaka DC2)

NetBox: Device List (23 devices registered)

NetBox: Location Management

Shumoku: Network Topology Visualization

Impressions: NetBox + Shumoku Handled all the tedious data entry — even wrote its own registration scripts Built a convincing demo environment from vague instructions Found and installed device templates from GitHub on its own Realized Shumoku needed cable data, then autonomously added and corrected the missing entries

03 containerlab Network emulation environment

containerlab — Building a BGP Learning Lab What is containerlab? A network emulation tool for building virtual router/switch topologies using Docker containers from a topology definition file. Supports vendor Docker images. Great for learning and testing. Instructions given 👤 Set up a containerlab environment 👤 Create a BGP learning lab 👤weUsedon't FRR for the routers — have much memory

containerlab BGP Lab — Claude Code's console output

Impressions: containerlab BGP Lab From vague instructions, auto-designed and built a 3-node eBGP lab (AS 65001–65003) Generated usage guides and learning materials automatically Configured FRR (a real network OS) without any help Even worked like a real engineer: created config files locally, then rsync'd them to the platform

04 OpenStack (DevStack) "Usually doesn't work properly" — IaaS platform

OpenStack (DevStack) Installation What is OpenStack? An open-source IaaS platform that manages servers, networks, and storage via API — like running your own public cloud. DevStack is a single-machine setup for learning and testing. Installation is hard. Configuration is hard. And it usually doesn't work properly! Instructions given H Set up an OpenStack environment on a single VM H DevStack works? Sure, go with that then

OpenStack Horizon: Dashboard (It actually worked!)

OpenStack: VM Console (test-vm booted)

Impressions: OpenStack (DevStack) DevStack — which "usually doesn't work" — actually installed successfully! Got stuck just like a human would, but checked logs and resolved issues on its own The broken VM console was also fixed autonomously Even wrote documentation — arguably more diligent than a human

Summary 🤖 🔄 📝 🚀 Claude Code is impressive 4 complex systems built with natural language alone. Faster than a human. Autonomous problem solving When it gets stuck, it mostly notices and fixes itself. Log checking and debugging are automatic. Human role: instruct and verify We're entering an era where you can get results without knowing the technical details. Even more powerful with customization This used bare Claude Code — add harness engineering and loops for even greater capability.

Frequently Asked Questions Isn't this dangerous? Q → Work on harness engineering. Building control mechanisms to prevent runaway behavior is key. Sounds complicated… Q → Just using bare Claude Code is surprisingly easy. Install it and set up SSH access — that's it. Worried about billing? Q → Usage is lower than you'd expect for bare usage. A Pro Plan is plenty for work like this.

What Does the Future Look Like? After harness comes the loop!! Things are moving fast and it's exciting. Add control and monitoring around Claude Code and you get even more autonomous operations. Feels like AI is actually increasing my workload… • Reminds me of when compilers got faster • All those auto-generated docs and systems still need a human to check them • Even if you build a checking system, a human has to check that too • Work doesn't get easier → jobs probably aren't disappearing Let's enjoy the ride and keep going!

That's a wrap! Let's continue at the networking party Ken Sasaki · JANOG57.5 Interim Meeting · 2026-06-19 🍺