Built-in agent

The agent in con is not the product. The terminal is.

The agent panel is there when terminal context matters and a model can help. It reads the pane you are using, acts in view, and asks before doing work that should not happen silently.

When to use it

Use the agent when the answer depends on terminal state:

• explain an error on screen
• plan a fix before changing files
• run a command after checking the current pane
• work inside an SSH or tmux session
• reason about a TUI or coding-agent CLI already running in a pane
• summarize what happened in a long terminal flow

Use the shell directly when you already know the command.

Provider and model choices live in Settings. Each tab keeps its own active provider/model once you choose it. Global settings define the default for new sessions and the models available in the picker.

What terminal-native means

The agent starts from the focused pane. From there it can reason about terminal objects instead of loose screenshots:

• visible pane output
• pane names and working directories
• SSH session context
• tmux sessions, windows, and panes
• shell state and command history
• TUIs and coding-agent CLIs running inside the terminal

That context is useful, but it is not a license to act silently. Destructive or high-impact actions still require approval.

When the agent streams markdown, con renders code, tables, math, and diagrams in the panel instead of treating the answer as a separate web page. Long responses stay scrollable so the terminal remains usable.

SSH, tmux, and TUIs

con is built for terminal-native workflows. The agent can help while you are inside SSH, tmux, shells, editors, and coding-agent CLIs.

Keep the same rule in mind: the terminal remains the source of truth. If the agent needs to know what is happening, it should inspect the pane before making claims or taking action.

Stay in control

• Keep the pane you care about focused before asking.
• Ask for a plan first when the task is broad.
• Review commands before approving them.
• Use the terminal directly for simple commands.
• Hide the agent panel when you want a plain terminal.

con should feel like a serious terminal with help available, not a chat app wrapped around a shell.

External agents

If you are building an orchestrator or subagent workflow, use con-cli and surfaces. Surfaces let another agent create worker terminal sessions inside a pane without taking over the main terminal layout.

The built-in agent harness and benchmark loop are open in the repository. They exist so terminal-native behavior can be tested and improved, not hidden behind a product claim.

If you want to repeat a workflow through the built-in agent itself, write it as a skill. Skills are the user-facing way to keep a good terminal routine without turning it into a separate app.