Chat and terminal modes

Every codeout session runs in one of two modes. Chat mode is a clean messaging UI over your agent, with tool approvals you tap and slash commands you type. Terminal mode is the raw PTY, for everything a real shell can do. You pick the mode when you open the session.

Chat mode

Chat mode is the headline. Instead of staring at a terminal on a phone screen, you get a proper conversation with your agent. The daemon does all the agent-specific parsing on its side and streams clean, typed events to your device, so the UI is the same whether you are on the web app or a native app.

Streamed replies and markdown

The agent's prose arrives token by token and renders as proper markdown: headings, lists, fenced code, the lot.

Thinking & reasoning streams

Live reasoning tokens stream into collapsible thinking blocks. Watch the agent plan and deduce in real time before it writes code.

Collapsible tool cards

Every tool the agent runs becomes a card you can expand: what it called, the input, and the result. Subagents nest inside the task that spawned them.

Tool-permission approvals

The agent asks before it runs anything. You get Allow once, Allow for session, or Deny, right in the chat, and the agent waits for your tap.

Dynamic slash commands

Type / for an autocomplete. /model switches models based on server allowlists, /effort tunes reasoning depth, /mode toggles chat vs terminal, and /clear starts fresh.

File and image uploads

Attach files and images straight from the composer, up to 100 MB each. They land in a folder on the daemon, never on someone else's server, and the agent can read them back without a permission prompt.

Sending files and images

Tap the attach button to pick files or images. They queue in a tray above the composer, so you can stack a few up and send them with your message. Each file can be up to 100 MB.

Uploads go to a folder on the daemon host, not into the cloud. They never leave your machine. The default is ~/.codeout/uploads, and you can point it somewhere else in Settings; the daemon hard-validates the path so it always stays under your home directory. The agent gets the file paths with your message, and because you uploaded them yourself, reading them back is auto-allowed, no Allow tap needed. Images open in a full-screen lightbox; tap any file to download it back to the device you are on.

The permission gate

This is the part that makes driving an agent from your phone sane. Tools do not auto-run. When the agent wants to use one, the daemon pauses it and sends you a prompt; nothing happens until you answer.

Allow once
run this one call, then keep asking. The safe default.
Allow for session
run it and stop prompting for that tool until the session ends.
Deny
the tool does not run; the agent gets a denied result and carries on without it.

Obvious read-only calls are auto-allowed by the agent's own classifier, so you are not tapping Allow for every file read. Anything that would have prompted in a normal terminal reaches the gate instead. If a turn ends with a prompt still open, the daemon denies it rather than leaving the agent hung.

Stopping a reply

While the agent is replying, a dedicated Stop button appears next to the composer. Tap it and the daemon aborts the turn mid-stream; the session stays alive and the conversation survives, so this is a course correction, not an ending. You can type your redirect while it stops and send it the moment the turn is gone.

Find in chat

The magnifier in the chat header searches the whole conversation: your messages, the agent's replies, and the tool output inside the cards. Matches are highlighted in place, and you jump between them from the find bar, so a file path buried in a tool card an hour ago is a search away instead of a scroll away.

The status bar

A chat session carries a status bar across the top, recomputed live from the agent stream, so you always know what you are talking to and how much room is left.

model
the model actually running. Tap the chip or use /model to choose from your server-configured model allowlist.
context window
how many tokens the model's window holds.
effort
the current reasoning effort, set dynamically by /effort.
context fill
how full the window is right now, so you can see a long session creeping toward the edge.
rate-limit windows
the agent's own usage windows and when they reset, straight from the stream.

Terminal mode

Sometimes you do not want a chat, you want the actual terminal. Terminal mode gives you the raw pseudo-terminal: full-screen TUIs, colours, cursor positioning, and anything else that works in your shell. The agent, or a plain shell, runs exactly as it would on the machine itself, and your device renders the live frame.

Pick terminal mode when you need a TUI the chat UI cannot express, when you want to run something that is not an agent at all, or when you simply prefer the keys-and-bytes feel of a real terminal. Both modes ride the same encrypted channel and the same paired devices; the only difference is what the daemon streams and what the device draws.

Which mode to use

Use chat mode for

Day-to-day agent work from a phone: reading replies, approving commands, switching model or effort, and keeping a readable history you can scroll.

Use terminal mode for

Full-screen TUIs, raw shell work, or anything that is not a chat with an agent. If it belongs in a terminal, it goes here.

Agents

Chat mode drives Claude, Codex, Gemini, and OpenRouter (with full tool execution), and you pick the agent per session. A plain shell works in either mode and needs no agent at all.