Async Mobile Dev: Six Claude Agents from a Phone

Read Articleadded Jan 4, 2026

The author runs six Claude Code agents on a Tailscale-only Vultr VM, controlled entirely from an iPhone using Termius, mosh, and tmux. A PreToolUse hook posts AskUserQuestion prompts to a Poke webhook, sending push notifications that enable asynchronous, interruption-friendly development. Git worktrees, deterministic port allocation, and cost- and security-conscious VM operations make the setup fast, parallel, and safe.

Key Points

  • On-demand, Tailscale-only Vultr VM (no public SSH) provides a secure, pay-per-use dev box controlled from an iPhone.
  • Mosh plus tmux delivers resilient, persistent terminal sessions suited to mobile usage; agent forwarding limitation is handled by falling back to SSH inside tmux for GitHub ops.
  • Push notifications via a Claude PreToolUse hook and Poke webhook alert the user when Claude needs input, enabling true async development.
  • Parallel work via Git worktrees and per-branch deterministic port mapping lets six Claude Code agents run simultaneously without conflicts.
  • A permissive but isolated trust model with tight cost bounds makes the setup safe and economical for iterative development.

Sentiment

Overall, the sentiment of the Hacker News discussion is mixed but leans towards appreciative of the technical ingenuity. There's considerable enthusiasm for the sophisticated setup and its productivity benefits, with many users sharing similar positive experiences and tools. However, a strong undercurrent of apprehension exists regarding the practical challenges of deep work on mobile and, more significantly, the societal implications of an 'always-on' work paradigm it might enable, leading to concerns about work-life balance and burnout. The community is engaged and divided between technological excitement and socio-ethical concerns.

In Agreement

  • Several commenters appreciate the sophistication of the setup, particularly the use of Tailscale, and share similar experiences using it for remote development.
  • The use of Git worktrees with Claude is highly praised as an effective way to manage multiple features or tasks concurrently, often combined with tools like tmux.
  • Users confirm the utility of custom notification hooks (e.g., via Home Assistant or similar) to be alerted when an AI agent requires input, enabling an asynchronous workflow.
  • Some users already employ similar setups (e.g., Tailscale + Terminus + home machine, or ConnectBot + Gemini CLI) for remote AI-assisted development, validating the general approach.
  • Anthropic's integrated mobile 'teleport' sessions for Claude Code are mentioned as a smooth alternative for mobile AI interaction, suggesting the market is moving in this direction.
  • The idea of 'cycling through agents, each working on a task, checking their work, unblocking them' resonates with how modern AI-assisted development feels to some.

Opposed

  • Concerns are raised about the practical limitations of mobile interfaces for deep work, such as typing long messages, reviewing diffs, or performing thorough code verification, which often necessitates a laptop.
  • Some users find the constant context switching across multiple parallel tasks, especially when interspersed with real-world activities, to be inefficient and lead to needing to re-read history extensively.
  • A significant ethical and societal concern is voiced about the potential for such always-on mobile workflows to lead to a 24/7 work culture for white-collar workers, eroding work-life balance and fostering 'meaningless products'.
  • One commenter suggests that if code cannot be run and verified on the mobile device, it limits how long Claude can operate autonomously before a laptop is needed, akin to 'driving blind'.
  • Security concerns are briefly mentioned regarding Claude potentially misusing a valid SSH key on a disposable VM, despite the author's stated defense-in-depth approach.
  • The 'lack of planning mode' in Claude Code web (compared to CLI) is a point of contention for some, impacting their preference for how to interact with the AI.