Configuring Cursor, Claude Code, Codex, and Aider one by one is painful: the browser works, terminals and IDEs still time out. Tun captures more traffic via a virtual adapter—great for an AI toolkit, not automatically “faster,” and it fights corporate VPNs if both stay on.
Three paths: system proxy (browser ChatGPT often enough), Tun (CLI/Electron IDEs), or app/env proxies when Tun isn’t allowed. Basics: Tun; start failures: Tun start failed.
Pick the path by situation
| Situation | Use |
|---|---|
| Browser ChatGPT is enough | System proxy |
| Terminals / Electron IDEs time out | Tun mode |
| Tun not allowed (managed machine) | App / env proxy per tool |
Recommended five steps
- Working nodes; latency not all red (latency).
- Disable other VPNs/boosters; install service mode if needed.
- Enable Tun; connections should show the target tool.
- Verify browser first, then CLI—don’t reverse that.
- Still failing: swap region nodes, don’t only flip ports (ChatGPT, Cursor, Claude Code, Codex).
Corp VPN and shutdown
Time-box office VPN vs Tun; don’t dual-stack forever. DIRECT for LAN (DIRECT rules). End of day: disable Tun, keep/clear system proxy as you prefer, glance at connections.
Tun on but AI still fails
Check node region, latency, DNS (DNS, slow speed). Tool-specific pages cover Gemini/Copilot/Aider. Rule haircuts: routing.
Split from system proxy
Web browsing: system proxy is often enough. IDE/CLI: prefer Tun over typing ports everywhere (ports). Builds: download center.