Agents
A pane can hold a chat with an agent, such as Claude Code, instead of a shell. You type a message in the box along the bottom, the agent answers above it, and it works on the files in the pane's directory as it would in its own program.
Your messages are bubbles on the right. The agent's answers run the width of a column down the middle of the pane, with Markdown laid out as it's meant to look: headings, lists, quotes, links, and code in the terminal's font on a background of its own.
Opening a chat #
| Does | |
|---|---|
Ctrl+B a | Opens a chat to the right of the focused pane. |
:Agent | Opens a chat in a new tab. |
:|Agent, :-Agent… | Opens a chat beside the focused pane, as `: |
A chat works in the directory a new terminal would start in: the one the pane it splits is in, or for a new tab, the one the workspace's focused pane is in. That's where the agent reads and changes files, and runs its commands.
Like any new pane, a chat opens in normal mode: i starts typing into it. The first chat starts the agent, which takes a second or two. Until then the chat says it's starting, and a message sent meanwhile waits and goes when it can.
Typing #
In insert mode, what you type goes in the box at the bottom.
| Key | Does |
|---|---|
| Enter | Sends the message. |
| Shift+Enter | Starts a new line in the message. |
| Ctrl+C | Copies what's selected in the message, or else, while the agent answers, asks it to stop. |
| Backspace, Delete | Delete a character, or with Ctrl, a word. |
| Left, Right | Move by a character, or with Ctrl, a word. With Shift, they select. |
| Up, Down, Home, End | Move by a line, or to its start or end. With Shift, they select. |
| Ctrl+A | Selects the whole message. |
| Shift+Tab | Moves the agent on to its next mode. |
| Esc | Normal mode. There's no program in a chat to need Esc. |
Clicking in the box puts the cursor there, dragging selects, and a double or triple click selects a word or a line. A paste goes into the message whole, lines and all, without sending it. Text can be typed through an input method: what's being composed shows in the box, at the cursor, until it's committed.
A message beginning with / goes to the agent as typed, which is how agents take their own commands, such as Claude Code's /compact.
One message is answered at a time. While the agent answers, the button in the box's corner is a stop button, and Enter keeps what you typed until it's done. Clicking the button with nothing being answered sends the message, as Enter does.
What you've typed but not sent belongs to the window: another window, or this one after it closes, starts with an empty box.
The agent's settings #
Along the bottom of the box are the agent's settings for the chat, as the agent offers them. For Claude Code:
- On the left, its mode: how much it does without asking. Manual asks before every change, Accept edits makes edits without asking, Plan plans before it acts, and Auto lets Claude decide what needs asking about. A new chat starts in Auto.
- On the right, its model and effort, such as
Sonnet 5.5 · High. - Beside them, how much of the model's context the chat fills, such as
38k / 200k.
Click a setting to open its menu, above the box. Up and Down move through it, Enter or a click picks, and Esc or a click elsewhere closes it. Shift+Tab steps through the modes without opening the menu.
The model and effort share a menu:
- Models by family: Opus, Fable, Sonnet and Haiku, each with the version it picks and what it's for, and the one Claude Code recommends marked. Picking a family picks its newest version, and keeps up as new ones come out.
- A family's versions: the
▸button at the end of its row, or Right, shows its other versions, to pin the chat to one. Left hides them. A chat on a pinned version opens the menu with that version shown and ticked. - Effort on a slider, from Low to Max: click or drag along it, or with the slider picked out, press Left and Right. It stays open as you slide, so you can try a few. Default, beside it, leaves effort to Claude Code, which picks it for each model.
Models from another agent are grouped only when one choice in a family always means its newest version, as Claude Code's opus does. Otherwise they're listed one per row.
A chat keeps the settings you pick. If the agent stops, or the daemon is upgraded, the chat asks for them again as its session reopens.
Normal mode #
In normal mode, keys move the view instead:
| Key | Does |
|---|---|
j k, Down Up | A line down or up. |
| Ctrl+D Ctrl+U, Page Down Page Up | Half a view down or up. |
gg G | The start or end of the chat. |
i, a, Enter | Back to insert mode. |
The mouse wheel scrolls in either mode. While you're at the end, the view follows the answer as it grows. Scrolled up, it stays put, and the bottom bar says so.
What the chat shows #
Below each message you send, what the agent did with it, in order:
- its words
- its thinking, if it shows any, quieter and slanted
- each tool it uses, on a line of its own:
○while it runs, then✓once it has finished, or✗if it failed - each command it runs and each change it makes to a file, on a card of its own
A command's card has the command at its head, after $. While the command runs, the card says what it's for. Once it's done, it shows the last 8 lines of its output, which is where what went wrong usually is.
A change's card has what the agent did at its head, such as Edit Cargo.toml, and how many lines it adds and removes, such as +3 −1. Under that come the lines, with those added on green after + and those removed on red after −. A new file is all added. A card shows the part of the file changed, not the whole file. Once the change is made, Claude Code also gives 3 lines either side of it, and where in the file it is.
Lines are numbered by where they are in the file: added and unchanged lines by where they are now, and removed lines by where they were. Before the change is made, as while the agent asks to make it, an edit's lines have no numbers yet. A new file is numbered from 1.
Folded, a card shows its first 8 lines (for a command's output, its last 8), each cut off at the card's edge. Click a card with more to show to open it: its lines wrap, to be read whole. Click again to fold it. Whether a card is open belongs to the window.
The daemon keeps the last 32 KB of a command's output, and the first 256 KB of each side of a change.
While the agent works, the turn ends with Working…. A turn the agent didn't finish says why: Stopped., or what went wrong. The tab is named after the title the agent gives the conversation.
When the agent asks #
In Manual mode, before it does anything that changes something, such as writing a file or running a command that isn't only reading, the agent asks. Other modes ask about less, and Auto leaves it to Claude. Its question takes the box's place, with an answer on each button: the first way to say yes in the accent colour, and the ways to say no quieter.
Click an answer, or press its number. Esc says no. Ctrl+C says no and stops the answer.
What it asks about follows the agent's own settings. Claude Code reads its usual settings files, so commands you've allowed in Claude Code are allowed here too.
Which agent #
Chats talk to Claude Code. clackatoo speaks ACP, the Agent Client Protocol, which Claude Code doesn't speak itself, so the daemon runs Claude Code's ACP adapter, @agentclientprotocol/claude-agent-acp, with npx. That needs Node.js, and the first chat downloads the adapter. The adapter runs a copy of Claude Code that comes with it, at a version tried with clackatoo, not the claude you may have installed.
It uses the account Claude Code is logged in to. Log in once with claude in a terminal, and every chat uses that login. With a Claude subscription, chats count towards the subscription's limits, as Claude Code does.
To talk to another agent that speaks ACP, set CLACKATOO_AGENT for the daemon to its command, split at spaces. The daemon reads it when it starts.
One agent process serves every chat. Claude Code's adapter starts a Claude Code process for each chat as well, of about 250 MB, and the adapter itself takes about 115 MB. Closing a chat's pane ends its process, and closing the last chat stops the agent.
What lasts #
The daemon keeps chats, as it keeps terminals. Closing the window closes nothing: a chat keeps answering, and the next window shows it.
Upgrading the daemon keeps each chat's conversation, but stops the agent. An answer that was under way is cut off, and the turn says so. The next message starts the agent again and carries on the same conversation. A chat from a build before chats were kept as turns comes through as one block of its text. Tools from a build before cards come through as lines, without their output or changes. Changes from a build before line numbers come through without them.
If the agent stops on its own, the chat says so, and the next message starts it again, carrying on the same conversation.
When the daemon itself stops, its chats go with its terminals.
clackatoo capture prints a chat as text: each message after a >, and what the agent said.
Not yet #
- Text in the conversation can't be selected or copied yet: only the message being typed.
- The thinking and tool calls of a turn don't fold away under a "Worked for…" line yet.
- Commands the agent runs show their output once they're done, not as they run. They'll run in terminals beside the chat, which you can watch and take over.
- Claude's 5-hour and weekly usage limits aren't shown.
- Images can't be pasted into a message.
/commands aren't listed or completed as you type them.