3.6.0 brings AI coding agents into Konsoll: Claude Code, Codex, Qwen Code and opencode show what they are doing in the tab, ask for you through macOS notifications, and can use your terminal as context. Everything below is new since 3.5.0 (it was tried in five betas).

AI agents in your tabs

  • The tab shows what the agent is doing. The agents report their state from their own hooks: working, waiting for you (a permission or a question, shown in the tab’s tooltip), finished or failed. A tab with several panes shows the most urgent one.
  • Notifications when an agent needs you. When an agent waits for you, finishes or fails in a pane you are not looking at, macOS shows a notification; clicking it brings you to that pane. The “finished” notification names the task. Turn them off in Settings ▸ AI Agents.
  • Answer permission prompts without leaving what you are doing. When an agent asks to run a tool, the notification shows what it wants — the command, the file — with Allow and Deny. The answer goes to the agent’s pane as if typed there, and Konsoll stays in the background.
  • Window ▸ AI Agents (⇧⌘A) lists the panes running an agent across every window, the ones waiting for you first, with the same Allow / Deny buttons, the task each agent is working on and how long it has waited. The Dock icon counts the agents waiting for you.

Your terminal as the agent’s context

  • Send to Agent: right-click a pane to paste the selection, or the last command with its exit status and output, into an agent’s pane as its next prompt. With more than one agent open, the menu lists each one by name, tab and task, the ones in the same tab first.
  • MCP server: konsoll --mcp gives an agent tools limited to the pane it runs in: open_pane (a split beside it running a command, without taking the focus), run_in_background and background_job_status (only the jobs that agent started), recent_commands (the commands run in its pane, with their output), list_panes and read_pane, which reads another pane — a dev server’s log, a test run — only after you allow it. Settings ▸ AI Agents has the command or configuration to register it with each agent.

Agents in parallel

  • Shell ▸ New Agent in Worktree… creates a git worktree of the current repository on a branch and starts the agent of your choice there in a new tab, apart from your checkout.

Agents over SSH

  • ssh host typed in a zsh or bash tab brings Konsoll’s shell integration to the host for that session, with nothing installed there. Claude Code, Codex and Qwen Code started on the host show their state in the tab, ask for permission through Konsoll’s notifications, and are listed by Send to Agent; the host’s commands become command blocks, so Send Last Command to Agent works there too.
  • The setup lives in a private temporary folder removed when the shell exits, and the agents get Konsoll’s hooks for that run only.
  • Logins with a remote command, -N, -W, a RemoteCommand in ssh_config or a host excluded in Settings ▸ AI Agents run ssh as usual, and so does command ssh. Turn it off in Settings ▸ AI Agents.

Connecting the agents

  • Settings ▸ AI Agents ▸ Connect adds Konsoll’s hooks to each agent’s settings file, keeping everything else (the previous file is saved as .konsoll-backup; a file with comments is left untouched), or installs the opencode plugin. Remove takes them out again, and Update appears when installed hooks are out of date.
  • A pane running an agent that is not connected offers to connect it.
  • Codex runs new hooks only after you trust them once with /hooks.
  • Claude Code can also use the Konsoll plugin from this repository’s marketplace instead.

Fixes

  • A crashed program no longer leaves mouse, focus or keyboard reporting on (every wheel turn used to reach the shell as text such as 65;67;25M).
  • An oh-my-zsh theme that declares its prompt variables with local no longer loses them (the prompt showed only the directory).
  • A command whose line wrapped right after a space no longer loses that space when sent to an agent or read with recent_commands.

Release v3.6.0 on GitHub