The shift toward managing AI coding agents directly from mobile devices is gaining traction among infrastructure-focused developers. In a recent deep dive, a developer detailed their migration from a custom solution named 'herdr' to T3 Code, an MIT-licensed agent harness control surface. The switch, documented on unbeholden.dev, highlights the growing demand for tools that allow engineers to spawn and monitor agents like Claude Code and Codex without being tethered to a desktop terminal.

The Architecture: One Server Per Container

The core of this workflow involves running a T3 Code server instance inside each dev container on an LXC host named Oren. This architecture ensures that agents remain sandboxed within their respective containers, preserving the security boundary that prevents permission prompts. The developer removed previous scripts like dcclaude and dccodex, replacing them with a unified dcup script that launches the T3 server using t3 start --no-browser --host 0.0.0.0 --port 3773. By binding the /history/t3 directory to the host, the setup ensures that thread data and pairings survive container rebuilds.

Mobile-First Workflow and In-Container Browsing

The primary appeal of T3 Code for this user is the ability to manage work from a phone via dedicated iOS and Android clients. Unlike previous setups that required SSH into a host and manually spawning terminals, the mobile UI allows for seamless resumption of agent tasks. A standout feature is the built-in browser, which runs inside the container itself. This eliminates the need to expose ports to the host for previewing changes, a significant simplification for developers juggling multiple projects that might otherwise conflict on standard development ports.

The Friction: Pairing and Early-Stage Bugs

Despite the benefits, the workflow comes with significant friction due to the project's 'very very early' status. The developer noted that t3 serve failed to correctly bootstrap projects, forcing a switch to t3 start. Pairing each container individually is tedious, requiring a custom dcpair script to generate QR codes with correct host addresses, as the default pairing links are often unreachable from mobile devices. Additionally, the Nix package includes Electron, resulting in a hefty 3.7 GiB footprint, and the lack of a centralized server means devices must be paired with every single container separately.

Key Takeaways

  • T3 Code enables mobile-first management of AI agents, supporting iOS, Android, and web clients.
  • The in-container browser removes the need for exposed host ports, simplifying network configuration.
  • Pairing is currently fragmented, requiring one-to-one connections for each dev container.
  • The tool is in early beta, with known issues around worktree integration and service persistence after host restarts.

The Bottom Line

T3 Code is a promising step toward decoupling agent management from the desktop, but its current fragmentation and heavy footprint make it a tool for those willing to script their way around the gaps.