CodonTerminal

The grid

• Add or remove panes with + and − in the toolbar — anywhere from 1 to 10.
• Click a pane to focus it; the highlighted border shows where your keyboard (and voice) goes.
• ↻ Restart relaunches every pane when you want a clean slate.
• Drag the boundary between panes to resize them (double-click a boundary to even them out), or pick a shape from the ▦ Layout menu — Auto grid, Two columns, Single row, Single column.
• ⌨ Compose docks a multiline input bar under the panes: Enter sends, Shift+Enter adds a newline, ↑/↓ recall history.
• Zoom with Ctrl + / Ctrl − / Ctrl 0, or Ctrl + mouse wheel.

One task per pane

Every pane is an independent agent session. The workflow that makes CodonTerminal click: give each pane its own job — one refactors, one writes tests, one explains an error, one drafts docs — and walk the room like a reviewer instead of waiting on a single chat.

Models per pane

• Each pane picks its own model — run Opus on the hard problem and a faster model on the busywork.
• Anthropic's Claude models are built in; hundreds more are available through OpenRouter or any OpenAI- or Anthropic-compatible endpoint. With the Codex CLI installed, its models appear too.
• Switching a pane's model keeps the conversation: swaps between Claude models happen live with scrollback intact, and a swap across providers carries the conversation over. Ctrl+Shift+M opens a type-to-filter quick-switch for the focused pane.
• The usage pill in the toolbar tracks what you're spending — your OpenRouter balance alongside today's tokens per agent, with the full breakdown on hover. Click it to run the live plan-limit check in a pane.

The task board

▤ Board opens a kanban board — To Do, In Progress, In Review, Complete — saved with your project, so it survives a restart. Each card carries a task, a priority, a role, and a model.

• Drag a card into In Progress and a pane picks it up: it opens on that card's model and starts on the task.
• ✨ Suggest asks your signed-in Claude to read a task, weigh it against the live OpenRouter catalogue and its prices, and pick the cheapest model that can actually do it — with a one-line reason. A model you don't have yet is added for you.
• 🚀 Run Mission takes one free-text goal and fills the board with an ordered set of cards, each with its own priority, role, and model.
• ⧉ Agents is a second tab showing the fleet at work: who's running, on which model and agent, in what role, busy or idle.

Letting it run itself

▶ Run Swarm works the board without you. It starts To-Do tasks in priority order up to a concurrency limit you set, opens a pane per task, moves each card along as it finishes, and backfills the free slot with the next one.

• Turn on the Reviewer gate and every finished task goes to a reviewer pane, which inspects the actual changes before the card can reach Complete.
• A pane you're already working in can join the swarm — press SWARM in its header and the board routes tasks into that session instead of spawning a throwaway pane. It only ever hands work to an idle pane.
• If a worker lands on Claude's “trust the files in this folder?” prompt, the board holds it out of the pool and names it in the status, then picks it back up on its own once you answer.

Voice input

• Click the 🎤 button to toggle the mic, or hold Alt for push-to-talk — your words are typed into the focused pane.
• The first time you use the mic, the speech model downloads and is cached — it's about 2.3 GB, so it's worth doing on a connection you don't pay by the megabyte. After that it's instant, and it never downloads again.
• Speech recognition runs entirely on your machine. Nothing you say is sent to a cloud service.

Plain terminals too

A pane doesn't have to be an AI. Panes fall back to your normal shell (PowerShell, bash, zsh), so the same grid can hold a dev server, a log tail, and three Claude sessions.