CodonTerminal

Find your way around

• The left rail switches between Agents, Tasks, Activity, Design and Settings (Ctrl 1–4, and Ctrl , for Settings). Command opens the Command Center in your browser.
• Ctrl K opens the command palette: search or run anything from one box.
• Ctrl Shift T starts a new agent. Ctrl + / Ctrl − / Ctrl 0 change the terminal font size.
• The project switcher at the top holds every folder you've added; each keeps its own queue and brief.

Agents side by side

• Run up to 16 terminals at once — any tool, any model, or a plain shell (PowerShell, bash) for a dev server or log tail.
• Lay them out as an auto grid, side by side, stacked, or focus one. Your agents are restored when you relaunch.
• Change a pane's model from its header. Your vendor's built-in models run on your subscription; API models you add run on your key.
• Compose, docked under the grid, sends a multiline message to one agent or to all of them: Enter sends, Shift+Enter adds a new line.
• Agent panes start in each tool's own auto mode, so they edit files and run commands in the project folder without stopping to ask. Each header shows Auto or Asks; click it to switch. Settings → Agents & routing → “Run without asking” sets the default.

Tasks and the queue

Tasks is one queue: Queued → Running → Review → Done. Add a task with a title and what done looks like, then press Run to start it now. Otherwise the queue runs on its own, several tasks at a time; Pause queue holds it. Every run is the vendor's own tool in non-interactive mode (claude -p, codex exec, gemini -p, cursor-agent -p, copilot -p), inside the task's own git worktree, and its output streams live into the app.

• The queue takes tasks in priority order and starts each on the best available subscription, up to the parallel limit in Settings → Agents & routing (three at once by default, two per subscription).
• Order your subscriptions in Settings → Agents & routing. Auto tasks go to the first one with room; you can also pin a task to one tool or model.
• Agents in swarm runs may run commands unattended, but only inside the task's own worktree. Turn off “Let agents run commands unattended” in Settings → Agents & routing and they may edit files but not run commands.
• Each project has one brief (.codon/brief.md) that every task sees — the project's goal, conventions and how to test. It is tracked in git.

When a plan hits its limit

When a run hits a usage or rate limit, that subscription rests until its reset time — CodonTerminal shows when it's back — and the task goes back in the queue for the next one. A limit is not counted as a failure, and the work isn't lost: the next attempt picks up on the same branch.

Review, then merge

When a run finishes, its branch is committed and the task moves to Review with the agent's summary, the live log and a full diff. Nothing touches your working tree until you decide.

• Merge brings the task's branch into whichever branch your project has checked out, then cleans up the worktree. If the merge would conflict, nothing changes and CodonTerminal tells you.
• Send back with a note: the next attempt continues on the same branch and sees your feedback.
• Discard deletes the branch and the worktree.
• Folders that aren't git repositories run in place, without a worktree — the task says so.

API models as overflow

Add OpenRouter, any OpenAI- or Anthropic-compatible endpoint, or a LiteLLM gateway in Settings → API models. They appear in every model menu, and Claude Code can drive them through CodonTerminal's built-in bridge. For queued tasks they are overflow only: used when you switch on “Overflow to API models”, and overflow stops for the day once today's metered spend reaches the daily cap you set. A task you pin to an API model, or an agent pane you run on one, isn't held by the cap. Keys are encrypted with your operating system's keychain and never leave the app's main process except to reach the provider.

Gattaca Plus: Plan and Second opinion

Everything above is in the one-time purchase. A Gattaca Plus key, added in Settings → License, adds two features. Plan turns one goal into a queue of tasks using a signed-in tool. Second opinion has a model from a different vendor review a change before you merge it. A checked key is trusted for a day, and the app keeps working offline for a week after the last good check.

Codon MCP and the Command Center

Settings → Connect tools gives you ready-to-paste commands that connect Claude Code, Codex, Cursor and other MCP clients to the same queue, using the MCP server bundled with the app. The Command Center shows the same queue, routes and live runs in your browser; it is served by the app on your own machine, and you sign in to it with a one-time link the app opens for you.

Design Mode

Open Design, enter your page address and pick an element in the live preview. CodonTerminal captures its HTML, the relevant CSS and a cropped screenshot. Review the capture for anything private, add your instructions, and turn it into a task or send it to an agent. The screenshot and page reference are attached as files, not pasted into the prompt, so your agent needs image support to look at the picture.

Activity

Activity keeps the latest 1,000 prompts your agents received, on this computer: typed into a terminal, sent from Compose, spoken, or handed over as a task, plus swarm runs. A Right now card shows what each agent is working on. Search them, reuse one in Compose, copy or delete it, even after the agent has closed. Shell panes aren't recorded, and anything that looks like a secret is scrubbed.

Voice input

• Voice is on Windows for now. Hold Alt to talk to the focused agent, or click the mic to dictate into Compose: your words appear at each pause, ready to check and send.
• The first time you use the mic, the speech model downloads and is cached — about 2.3 GB, so do it on a connection you don't pay for by the megabyte. After that it starts straight away.
• Speech recognition runs on your machine. What you say isn't sent to a speech service.