Driving your coding agents from a phone: the control surfaces compared
By Sheela ·
The coding agent is no longer the scarce part. Most developers now hold two or three subscriptions, and the friction has moved to running them: one terminal at a time, tethered to a laptop, with no good way to watch a long job or juggle several at once. A layer of open tools has grown up to fix exactly that, and none of them is a model.
The remote surface
T3 Code (4.2), from Theo Browne's ping.gg, calls itself an agent harness control surface. It drives whichever agent CLIs you have signed in to locally, Claude Code, Codex, Cursor, Grok Build, OpenCode and Google Antigravity, and exposes them through real iOS and Android apps, a web app and a desktop build. Start a run at your desk, approve a step from the bus. It is MIT-licensed and, with about 8,600 forks against 32,600 stars, unusually serious about letting you fork it. The caveats are its age, a v0.1.x release with thousands of open issues, and that opening a server to your phone is a remote-access decision you own.
The parallel surface
A second group is built around running many agents at once. Claude Squad (3.9) manages multiple agents in isolated git worktrees so they do not trip over each other, and Vibe Kanban (3.4) puts agent tasks on a board you drag between columns. Both answer the same question, how do I keep five agents busy without five terminals, from opposite directions: one terminal-native, one visual. Vibe Kanban is the weaker of the two in our reading, with a thinner track record.
The all-in-one
Then there is OpenCode (4.6), which blurs the line because it is both an agent and, increasingly, a surface: a terminal-first client that T3 Code can also drive. At about 222,000 stars it is the most adopted open agent here, and for a developer who wants one well-built TUI rather than a control layer over several back ends, it is the simpler answer.
What they share, and where they bite
None of these tools sandboxes anything. They inherit the permissions of the agent CLIs underneath, which run shell commands and edit files as you. A control surface adds reach, not safety, so the security story is still the agent's own approval prompts and whatever isolation you add. They also all move fast and version low; if you build a workflow on one, pin a release.
Which to use
If the pain is being chained to a laptop while an agent works, T3 Code's phone app is the only one that solves it directly. If the pain is running several agents without collisions, Claude Squad's worktrees are the cleaner mechanism. And if you would rather own one excellent client than a layer over many, OpenCode is where to start, and still works under T3 Code later if you change your mind.