Claude Code On-The-Go demonstrates six agents running from a phone via a private VPN and cloud VM, proving mobile AI development is practical without a laptop.
AI Team
Continue your reading
Claude Code On-The-Go from Anthropic Claude demonstrates mobile AI development in the wild.
The story, published by mgranda on Granda, shows six Claude Code agents running in parallel from a phone with no laptop or desktop in sight. The setup centers on Termius on iOS for remote access, a cloud VM, and a private network built around a VPN so the work can continue without tethering to a traditional workstation. This isn’t just a curiosity; it foreshadows how developers will increasingly ship AI copilots and agents from anywhere, not just from a developer desk.
The hardware is straightforward but precise. A Vultr VM in Silicon Valley holds the workload, specifically instance vhf-8c-32gb, with a documented cost of $0.29/hour (~$7/day when running). Access is intentionally restricted to a private network, emphasizing defense in depth rather than exposing a public SSH endpoint. Two scripts handle the lifecycle: vm-start to boot and wait for network readiness, and vm-stop to shut the machine down when work is done. An iOS Shortcut is used to kick off the VM via the Vultr API before the user even opens Termius, delivering a one-tap start to compute.
The network design leans on a private mesh rather than open ports. The VM sits behind a cloud firewall that blocks everything except the Tailscale coordination layer, with local nftables as a second line of defense and fail2ban to deter brute force attempts. This is not about a fancy desktop IDE connected to the cloud; it is a small, repeatable stack that makes mobile development practical. The explicit absence of a public SSH surface reduces exposure and keeps the control plane consolidated through a private network.
Mobility is where the approach shines. The author relies on Mosh to keep the session alive as the phone moves between WiFi and cellular networks, or even when the device goes to sleep. The workflow hinges on a persistent SSH session tunneled through Mosh, using a command like mosh --ssh = "ssh -p 47892" [email protected]. One gotcha is that Mosh does not forward SSH agents, so Git operations requiring GitHub authentication are performed inside a tmux session using regular SSH keys. This is a pragmatic reminder that mobile remote development still requires attention to classic SSH workflows.
What this means for developers and the tooling stack is meaningful. It decouples AI agent development from the local workstation and demonstrates a mobile-first model for async coding. This sits alongside other AI-assisted workflows like GitHub Copilot or cloud-based IDEs; Claude Code On-The-Go shows how remote compute and a private network can keep AI agents fed during real-time, on-the-go work without sacrificing security or control. It also underscores the value of reliable, mobile-ready networking tooling as a core part of practical AI development stacks.
Going forward, this pattern will likely become more common as mobile devices acquire more capable compute, networking stacks, and automation hooks. Developers will increasingly pair private cloud compute with lightweight frontends on phones, trading the traditional laptop for a compact, secure remote workstation. The real test will be managing costs at scale, keeping latency acceptable for interactive AI sessions, and maintaining secure, auditable pipelines when agents run from distant clouds. If anything, Claude Code On-The-Go points to a future where mobile devices serve as the handier control plane for AI-assisted development, not just a screen for reading prompts.
Claude remains the centerpiece, but the real story is the stack that makes mobile AI development feasible: a private VPN, a cloud VM, and a lightweight orchestration that keeps six agents fed from a pocket-sized device. For teams building on top of this approach, expect more emphasis on secure access patterns, more automation for VM lifecycle, and more guidance on how to balance on-device prompts with remote compute. The era of truly portable AI development is not a distant dream; it is a practical workflow already in use by ambitious developers.
< a href="https://granda.org/en/2026/01/02/claude-code-on-the-go/" target="_blank" rel="noopener"> Tailscale is the glue that keeps the private network cohesive, while Vultr provides the right balance of cost and proximity to the developer’s location. If you want to explore the underlying tooling, the MOSH project is the enabler for resilient mobile SSH sessions across network transitions.