Agents and scripting

Mightty works with coding agents in two ways. It notices when an agent running in a pane is waiting for you, and it lets agents and scripts control panes through an MCP server and a command-line client.

When an agent is waiting for you

Mightty recognizes these agents by their command name: claude, codex, gemini, aider, opencode, amp, goose, cursor-agent, copilot, crush, qwen and droid. It also recognizes them when they run as a script under node, bun, deno or python.

A pane needs your attention when any of these happens:

An agent in it has been working and then goes quiet.
A program in it rings the bell.
A program in it asks for a notification, with OSC 9, OSC 777 or mightty ctl notify.

If you aren't looking at that pane, Mightty tells you in three places:

The pane gets an orange border, and the title of its window starts with ●.
The Dock icon shows the number of panes that need you. It bounces once if Mightty is in the background.
A notification appears if you are in another app.

Press ⌘⇧A to go to the pane that has been waiting longest.

Claude Code

Claude Code can tell Mightty the moment it asks for permission, without waiting to go quiet. Add this hook to ~/.claude/settings.json:

{ "hooks": { "Notification": [{ "hooks": [{ "type": "command", "command": "mightty ctl notify 'Claude needs you'" }] }] } }

MCP

mightty mcp is an MCP server that gives an agent access to panes. To add it to Claude Code:

claude mcp add mightty -- mightty mcp

The server provides these tools:

Tool	What it does
`run_command`	Runs a command and returns its output and exit code.
`open_pane`	Opens a pane: a split beside the pane you are in, a tab, or a window.
`read_pane`	Reads a pane's screen or scrollback.
`list_commands`	Lists recent commands in a pane, with their output and exit codes.
`wait_for`	Waits for text to appear, for a command to finish, or for the pane's program to exit.
`send_text`, `send_keys`	Types into a pane.
`list_panes`, `focus_pane`, `close_pane`	Lists panes, brings one to the front, or closes one.
`notify_user`	Marks the agent's pane as needing you.

Some details of how the tools behave:

Panes the agent opens don't take keyboard focus, and they stay open after their command ends.
run_command without a pane runs in a hidden shell. The agent reuses that shell for the whole session, so cd and environment variables carry over from one call to the next. Mightty closes the shell when the session ends.
Access is limited to the agent's own pane and the panes it opened. To give an agent access to every pane, start the server with mightty mcp --all-panes.
Without a running Mightty, mightty mcp runs panes itself, with no windows.

mightty ctl

Every running Mightty has a control socket. mightty ctl is a command-line client for it:

mightty ctl list                           # panes: state, window, title, folder
mightty ctl run --pane 1 -- make test      # run in pane 1, print the output, exit with its code
mightty ctl spawn --right -- npm run dev   # a split beside this pane (or --down, --tab)
mightty ctl wait 3 --text "ready"          # wait until pane 3 shows "ready"
mightty ctl read 3 --lines 20              # the last 20 lines of pane 3
mightty ctl keys 3 ctrl-c                  # press keys in pane 3
mightty ctl attention                      # panes that need you, longest waiting first
mightty ctl watch command.finished         # events as JSON lines

Run mightty ctl with no arguments to list every command and flag.

Which Mightty it talks to. Inside a Mightty pane, mightty ctl uses the Mightty that pane belongs to. Anywhere else, it uses the one that was started most recently.

Commands after --. run and spawn take the command to run after --. Several arguments are passed as separate words, so -- rm "my file" removes one file. A single argument is passed to the shell as written, so -- 'make && ./run | less' works as a shell line. An argument that contains a control character is refused, because the terminal would act on it as a key press.

run without --pane starts a hidden shell for the command and closes it when the command finishes.

JSON-RPC

The socket speaks JSON-RPC, so a program in any language can use it directly. mightty ctl call METHOD [JSON] makes a single call and prints the result.