Tailscale + VPS: The Ultimate Network Setup for Remote AI Coding

The biggest pain point in AI development from home is not GPU power. It is network connectivity. HuggingFace times out. GitHub clones fail. Conda downloads at 10 KB/s.

I spent weeks figuring this out. If you have a VPS in a well-connected region, combined with Tailscale, you can solve almost all network problems. This is my journey and a complete guide.

The Pain: Network is the Silent Killer

You have experienced these scenarios:

• pip install transformers hangs on Resolving for minutes
• git clone a 5GB model repo, fails overnight
• OpenAI API calls timeout constantly
• HuggingFace .safetensors downloads reset mid-way

I have a decent Linux machine at home. Hardware is fine. Network is not. Then I got a Netherlands VPS (cheap) and noticed the same code runs flawlessly there.

The question became: how do I make my home machine "borrow" the Netherlands network?

Solution 1: Tailscale Exit Node (Recommended)

This is my current setup. The principle is simple: make the Netherlands VPS a Tailscale exit node, and route all traffic from the home machine through it.

To the outside world, your home machine appears to be in the Netherlands.

Step 1: Install Tailscale on Both Machines

Go to Tailscale, register an account, then install on both machines:

curl -fsSL https://tailscale.com/install.sh | sh

Login:

sudo tailscale up

A link will appear for browser authorization. After that, tailscale status should show both machines online.

Step 2: Configure Exit Node on the VPS

On your Netherlands VPS:

# Enable IP forwarding (required)
echo 'net.ipv4.ip_forward = 1' | sudo tee -a /etc/sysctl.d/99-tailscale.conf
echo 'net.ipv6.conf.all.forwarding = 1' | sudo tee -a /etc/sysctl.d/99-tailscale.conf
sudo sysctl -p /etc/sysctl.d/99-tailscale.conf

# Advertise as exit node
sudo tailscale up --advertise-exit-node

Then go to Tailscale Admin Console, find your VPS, click Edit route settings, and check Use as exit node.

Step 3: Use the Exit Node from Home

Assuming your VPS is named dutch-vps in Tailscale:

sudo tailscale up --exit-node=dutch-vps

Done. All your traffic now routes through the Netherlands.

Verify

curl ipinfo.io

If the returned IP and region show Netherlands, you are set.

Quick Toggle Aliases

Add these to your .bashrc:

# VPN on (route through VPS)
alias vpn-on='sudo tailscale up --exit-node=dutch-vps'

# VPN off (direct connection)
alias vpn-off='sudo tailscale up --exit-node='

Solution 2: HuggingFace Mirror (Fastest Downloads)

If your main issue is downloading large models from HuggingFace, use a mirror instead of a proxy. It is faster.

Add this to your terminal or .bashrc:

export HF_ENDPOINT=https://hf-mirror.com

Now huggingface-cli download and the transformers library will use the mirror automatically. Download speeds can max out your bandwidth.

Note: This only solves HuggingFace downloads. It does not help with OpenAI API or GitHub.

Solution 3: Precise Proxying (Proxychains)

If you do not want to route all traffic through the VPS, only specific scripts, use Proxychains.

Assuming you have a local proxy client (like Clash) on port 7890:

Method A: Temporary Environment Variables

export http_proxy=http://127.0.0.1:7890
export https_proxy=http://127.0.0.1:7890
python my_ai_script.py

Method B: Proxychains Force Takeover

Some libraries (especially those with C++ backends) ignore environment variables. Use Proxychains:

sudo apt install proxychains4

Edit /etc/proxychains4.conf, change the last line to:

http 127.0.0.1 7890

Then run:

proxychains4 python my_ai_script.py

Is Remote SSH Good for AI Coding?

This is another question I have been exploring. The answer: absolutely yes.

Why Remote Servers Work Better

1. 24/7 Uptime: Run AI tasks in tmux on the server. Close your laptop, it keeps running. Wake up to finished code.

2. Environment Isolation: AI-generated code often needs many dependencies. Install on server, not your local system.

3. Network Advantage: Overseas servers have more stable connections to OpenAI, Anthropic APIs.

4. VS Code Remote SSH: Local machine handles only UI. All code execution happens on the server. Feels like local development.

Recommended Tool Stack

Handling 10+ VS Code Windows?

Another problem I wrestled with: what if I need to work on ten or more Next.js projects simultaneously?

I tried four approaches:

Why 10+ Next.js Projects Drain Resources

Each Next.js (App Router) dev server consumes 500MB - 1.5GB memory. Ten projects means:

• Memory: At least 16-24GB just for Node.js, plus VS Code and browser. Start at 32GB, 64GB for stability.
• File watchers: Ten projects exhaust system file handles.

Fixing WSL2 Disconnects

If you use WSL2, the most common issue is random disconnects. Create .wslconfig in your Windows user directory:

[wsl2]
memory=48GB
processors=12
networkingMode=mirrored

networkingMode=mirrored is key. It fixes most random network disconnection issues.

Also, never put projects in /mnt/c/. They must be in WSL internal paths like /home/user/projects. This alone gives 10x speed improvement.

My Workflow Summary

After all this experimentation, my current workflow:

1. Daily development: Home Linux + Tailscale Exit Node (Netherlands VPS)
2. Model downloads: export HF_ENDPOINT=https://hf-mirror.com, mirrors max out bandwidth
3. Long-running tasks: SSH to VPS, run Aider or Claude Code in tmux, go to sleep
4. Multi-project parallel: Windows + WSL2, properly configured .wslconfig

Core principle: Make network problems disappear at the infrastructure level. Then you never think about it while coding.

Resources

• Tailscale Official
• Tailscale Exit Node Docs
• HuggingFace Mirror
• Aider - Terminal AI Coding Assistant
• tmux Complete Guide

If you are also working on remote AI development setups, feel free to reach out. This stuff has a learning curve, but once configured, it just works.