Verdict: For developers and builders in 2026, the local terminal is no longer the best home for your AI agents. By moving your workspace to a Virtual Private Server (VPS) and using Tmux, you can run agents like Claude Code 24/7 with full autonomy, bypassing the manual approval friction of native "Remote Control" features and ensuring your builds continue even when your laptop is closed.
Last verified: 2026-06-22 · Best for: Builders, SaaS founders, and Devs · Prerequisites: Basic SSH knowledge, VPS (DigitalOcean/EC2).
Why Claude's Native "Remote Control" Isn't Enough
In February 2026, Anthropic released Claude Code Remote Control, a bridge that lets you control your local terminal from the Claude mobile app. While convenient for quick checks, it has three fatal flaws for serious builders:
- The Sleep Problem: Your local machine must stay awake. If your laptop lid closes, the agent dies.
- Approval Friction: Remote Control mode often forces a manual "Accept" for every file write or bash command to prevent "dangerous" headless edits.
- Session Limits: Native remote sessions are currently capped at 32 per account, which is insufficient for developers managing large-scale multi-agent worker factories.
To achieve true autonomy, you need a setup that lives where the internet lives: in the cloud.
The "Always-On" Stack: VPS + Tmux + CLI Agents
This framework (often called the "Real Boys" setup by high-output builders) treats your AI agent as a remote employee that never sleeps.
| Component | Tool Recommendation (2026) | Price/Tier | Role |
|---|---|---|---|
| Cloud Server | DigitalOcean Droplet / AWS EC2 | ~$6-12/mo | The 24/7 execution environment. |
| Multiplexer | Tmux 3.6b | Free (Open Source) | Keeps the session alive after you disconnect. |
| AI Agent CLI | Claude Code v2.1 | Pro/Max Subscription | The primary builder (Opus 4.8 default). |
| Mobile Access | Terminus | Free / $10 (Pro) | Prompting your agent from your phone. |
| Code Review | Greptile (TREX) | Free Beta / $2 per run | Independent test/run validation of AI PRs. |
Step 1: Provision and Harden Your Workspace
Deploy a standard Linux VPS (Ubuntu 24.04+ recommended). Security is critical because your AI agent will have write access to your code and GitHub.
- Firewall (UFW): Disable all incoming traffic except for your SSH port.
- SSH Hardening: Disable root login (
PermitRootLogin no) and use Ed25519 keys only. - Port Change: Move SSH off port 22 to a high-numbered port (e.g., 2222) to dodge 90% of automated bot scans.
Step 2: Set Up the Persistent Session with Tmux
tmux is a terminal multiplexer that allows processes to run in the background. If your connection drops, the AI keeps working.
- Install:
sudo apt install tmux - Start a named session:
tmux new -s ai-builder - Run your agent:
claude
When you want to leave but let Claude keep building, press Ctrl+b then d to detach. To resume later, run tmux attach -t ai-builder.
Step 3: Enabling True Autonomy (Headless Mode)
The advantage of a VPS is that you can safely run Claude in "dangerous" mode if you trust your environment. Claude Code 2.1 supports a --dangerously-skip-permissions flag (or using Shift+Tab to toggle bypassPermissions in the TUI). This allows the agent to clone a repo, implement a feature, run tests, and send a PR while you are literally at the gym or asleep.
"Prompting from the Phone": The Mobile Workflow
With your VPS running tmux, you can use the Terminus iOS/Android app to connect to your server from anywhere.
- The Workflow: You get an idea during lunch, open Terminus on your phone, reattach to your
tmuxsession, and tell Claude: "Implement the Stripe webhook handler and send a PR." - The Result: You disconnect, and Claude begins the 15-minute process of writing code and verifying it via Greptile's TREX agent, which provides real-time execution traces and screenshots of the fix.
What this means for you: Information Gain vs. Commodity Chat
Building a persistent cloud workspace moves you from "chatting with AI" to "orchestrating AI." While most people are still copy-pasting code from a browser, you are running a unified Agent OS that builds, tests, and deploys autonomously. This setup is the foundation for owning and maintaining your AI agents rather than just using them.
FAQ
Q: Isn't running AI in bypass mode dangerous? A: Yes. Always run in a git-initialized directory so you can revert changes. Never give an agent access to your production database credentials unless you have a robust Sandboxing protocol in place.
Q: Which model should I use for long-running remote tasks? A: Claude Opus 4.8 is the 2026 standard for reasoning. However, for large codebase scans where you need a massive context window, GLM 5.2 review shows it is a powerful open-weights alternative with its 1M-token window.
Q: How much does this setup cost? A: A basic VPS is ~$10/mo. Claude Pro is ~$20/mo. Terminus is free for basic use. For under $35/mo, you have a professional-grade autonomous coding factory.
Q: Can I use this with Cursor?
A: Cursor is an IDE, not a CLI. While you can use "Remote-SSH" in VS Code/Cursor to connect to your VPS, it still requires your local IDE to be open. The CLI-first tmux approach is the only way to achieve 100% persistence without an open laptop.
Discussion
0 comments