Skip to content

Built-in terminal

Context Code embeds a real terminal (with a native pseudo-terminal, not an imitation): what runs there is your usual shell, inside the app and next to the agent chat.

  • Show/hide the terminal is available as a command in the quick search, in addition to the panel button.
  • You can split the view into several panes and keep, say, a server running in one and one-off commands in another.
  • The project’s run and build buttons dump their output here, in plain sight.
The terminal supports split panes: server on one side, commands on the other.

The agent runs its commands under the rules of the chat’s approval mode: in “Ask for approval” every command reaches you as a card before it runs; in “Read only” it runs nothing. Every command the agent executes is logged in the conversation along with its output.

Besides its own commands, the agent can look at the terminals you have open in the panel. Use it to ask “check what error the terminal shows” or “how is the server I have running doing”.

  • Reading needs no permission: the agent lists the open terminals and reads their last lines (80 by default, up to 400), without colors.
  • Typing into one of them, or stopping it with Ctrl+C, is treated as running a command: it always asks for approval, also in “Approve for me”, and the card shows exactly what will be typed.
  • It only sees terminals open inside the chat’s project; a chat from another project can’t read what happens in them.
  • What it reads first goes through the Vault masking: if a stored key shows up in the terminal, it reaches the model masked.

If you are connected to another machine through remote access (for example, a server running contextcode serve), the terminal you open in the panel runs on that machine, in the project folder. It can only be opened in projects registered there. The terminal stays on the machine where you opened it even if you switch connections later.