An AI Agent Controller for Claude Code: What to Buy and What to Bind

The Pedal Three is the plainest answer to what people mean by an AI agent controller for Claude Code: three ribbed foot pedals on a curved steel base, $69, which enumerate as a keyboard and send whatever you assign them. Bind accept, reject and push-to-talk, and the loop you run all day happens under the desk while your hands stay on the keys.

What counts as an agent controller?

Nothing formal, and no standard defines it. The term is doing a simple job: it names the small hardware you keep within reach because you now drive an agent hundreds of times a day and the actions are always the same few. In practice it has four parts, and almost nobody needs all four. A pedal or button for accept and reject. A pad for the shortcuts you repeat. A knob for scrolling long output. A light so you know when the run is over.

What does Claude Code specifically want bound?

Claude Code runs in the terminal and stops for you in two ways: it proposes an edit and waits for approval, and it sometimes pauses mid-run to ask a question. Those two pauses are most of your keyboard work. A sensible first mapping:

  • Accept the proposed edit, on the pedal or button you can hit without looking
  • Reject, immediately next to it
  • Interrupt, for the run that has clearly gone the wrong way
  • Push to talk, held while you dictate the next instruction
  • A knob to scroll, because agent output is long and the mouse is far away

Claude Code knows none of this is happening. You are pressing the same keys you already press, from a different object.

Which piece should I buy first?

Buy for the part of the loop that annoys you most, not for the full set.

  • Pedal Three, $69. Plug in the captive USB cable and it enumerates as a keyboard, so no driver is needed. Assign each pedal a key, a modifier combo, a text string or a mouse click with the manufacturer's utility.
  • KM16 Macropad, $119. Sixteen keys and three knobs for the people whose loop includes test, commit and pane navigation. It remaps in the browser through VIA, with no firmware flashing.
  • Status Light, $45. Powered over USB-C. Point a hook or a script at it: red when the task starts, amber when the agent stops to ask, green when it exits cleanly. Claude Code supports hooks, so this is the one part of the setup where the agent really can drive hardware, through a script you write.

What will it not do?

It will not read Claude Code's state on its own. Nothing here subscribes to a status feed, so a light knows only what your own hook tells it, and if you never write the hook the light sits there being a nice object. It will not improve your reviewing either. A bigger accept button makes rubber-stamping easier as well as approving, and that is worth thinking about before you make approval the most comfortable action on your desk.

Why a generic controller rather than a Claude Code one?

Because these send keystrokes, the same pedal works in Claude Code, Codex, Cursor, Copilot, Gemini CLI and Windsurf, and it still works when you change tools next quarter. The trade is genuine and we would rather say it out loud: a vendor-locked controller can carry the tool's own artwork and can be told directly what the tool is doing. This cannot. It is hardware that types what you bind it to, and you do the binding.

The light half of this is the part people underrate, so read our status light setup for Claude Code next.

Back to blog