Most AI automation content focuses on the tools you talk to — GPT, Claude, agents, prompts. Almost none of it covers the infrastructure layer underneath: the plumbing that keeps everything running when your AI setup spans multiple machines, lives on a remote VPS, and needs to stay operational 24/7.
We run a multi-machine AI environment — a VPS as the central hub, connected via Tailscale to several macOS machines (Mac minis and MacBook Pros). Our AI agents coordinate tasks across this network continuously. As the setup has grown, we’ve learned that infrastructure tooling matters as much as the AI itself.
This week we did a deliberate pass to fully equip the VPS. Here’s every tool that was installed and exactly why it earns its place.
1. tmux — The Session That Never Dies
If there’s one tool on this list that changes everything, it’s tmux.
When you SSH into a VPS and run a long process, closing that connection kills it. For AI workflows — indexing, batch processing, agent tasks that take 10 or 20 minutes — this is a serious problem.
tmux creates persistent terminal sessions that keep running even after you disconnect. You can detach, close your laptop, reopen hours later, and reconnect to find your session exactly where you left it. For background agent operations, it’s non-negotiable.
2. sshfs — Remote Files That Feel Local
sshfs mounts a remote machine’s filesystem over SSH so its files appear as if they’re local.
In practice this means the VPS can read and write files on a Mac mini directly — no FTP, no SCP commands, no intermediate copy steps. Agent configs, project files, generated outputs — all accessible across machines as if everything were one system.
3. rclone — Automated Backup for Everything
AI systems generate a lot of data: logs, generated content, agent memory files, conversation records. Without a backup strategy you’re one server failure away from losing months of work.
rclone connects to virtually any cloud storage provider and supports automated, scheduled transfers. Set it up once, forget about it. Critical agent data gets synced off the VPS to cloud storage on a schedule without any manual intervention.
4. nmap — Know Your Network
When multiple machines are talking to each other over a private network, visibility matters. nmap scans the Tailscale network and tells you exactly what’s running where — which services are active, which ports are open, which machines are reachable.
For debugging agent communication issues this is invaluable. Instead of guessing why one machine can’t reach another, you scan and know in seconds.
5. Tailscale CLI — The Backbone, Controlled
Tailscale is the VPN that stitches all the machines into a secure private network regardless of physical location. The Tailscale CLI gives command-line control — status checks, peer visibility, route management, quick diagnostics.
When an agent can’t reach a machine it should be reaching, the first thing you do is check Tailscale status. Having it in the terminal is faster and more scriptable than any GUI. (Already present — confirmed and verified.)
6. ncdu — Find Where Your Disk Went
A VPS with limited storage and AI agents generating logs, outputs, and temp files is a recipe for unexpected “disk full” errors at 2am. ncdu (NCurses Disk Usage) is an interactive terminal tool that shows exactly where disk space is going — drill down directory by directory until you find the culprit.
We’ve caught bloated log files, cached model outputs, and forgotten temp directories with this. Running it monthly is basic infrastructure hygiene.
7. syncthing — Real-Time Peer-to-Peer Sync
Unlike rclone (which runs on a schedule), syncthing syncs files continuously and in real time between devices — peer-to-peer, encrypted, and without routing through a central cloud provider. It works natively over Tailscale.
For shared configuration files that multiple machines need to stay current on — or agent memory that should be available across the network — syncthing keeps everything in sync automatically.
8. jq — Command-Line JSON Processing
Modern AI systems speak JSON. API responses, webhook payloads, agent outputs, structured logs — all JSON. Without jq, parsing any of it at the command line means piping through grep and awk and hoping for the best.
jq filters, transforms, and reformats JSON cleanly. When scripting automation that processes OpenClaw API responses or manipulates structured log data, jq is the tool that makes it readable and reliable.
9. yq — jq for YAML
If jq handles JSON, yq handles YAML — and YAML is the format of infrastructure. Docker Compose files, agent configuration files, CI/CD pipelines — all YAML. Being able to read, modify, and update YAML files from the command line without opening an editor is a major workflow accelerator.
For automating configuration updates across machines, yq is the tool that makes it scriptable.
10. htop + glances — System Health at a Glance
You cannot manage what you cannot see. htop gives an interactive, real-time view of every process running on the VPS — CPU, memory, load. glances extends this to include network I/O, disk activity, and per-process resource consumption in a single dashboard.
During intensive agent operations — large batch jobs, parallel AI calls, heavy indexing — these tools are how you catch resource bottlenecks before they become failures. They’re open in a tmux pane and checked any time something feels slow.
11. httpie — Human-Friendly HTTP Testing
httpie is what curl should be. It’s a command-line HTTP client that formats responses readably, handles authentication cleanly, and makes it easy to test endpoints without decoding a wall of text.
For testing OpenClaw API endpoints, verifying webhooks, debugging integrations between services — httpie turns a painful debugging session into a few clean commands. The output is readable by humans, not just parsers.
12. ripgrep (rg) — Search Everything, Fast
Codebases grow. Log directories accumulate. Configuration files multiply. ripgrep searches through all of it in milliseconds — recursive, regex-capable, gitignore-aware, and significantly faster than grep or find for large directories.
When debugging a problem that starts in a log file somewhere, or tracking down where a configuration value is set across dozens of files, ripgrep finds it instantly. It’s the search tool that makes large codebases navigable.
Why This Matters
None of these tools are glamorous. They don’t show up in AI demos. But collectively they turn a VPS into a capable, observable, resilient operations platform — one that can run agent workloads, stay in sync across machines, handle failures gracefully, and give you the visibility to know what’s actually happening at any moment.
The pattern here is intentional: persistent sessions (tmux), cross-machine file access (sshfs, syncthing), automated data protection (rclone), network visibility (nmap, Tailscale CLI), resource awareness (ncdu, htop, glances), and developer productivity (jq, yq, httpie, ripgrep).
That’s the full stack for running AI operations seriously. Not just prompts and models — but a real foundation under them.
We use this infrastructure to power AI-driven systems for local businesses. Get in touch if you’re interested in what that looks like applied to your operation.
