agentpane

agentpane keeps the subagents of a Claude Code session in a pane beside the transcript. Each row says which tool call the agent is running, how long it has run, its model and its tokens, and a click opens the agent’s whole conversation in the pane. It is for sessions where Claude fans work out to several agents and the footer list says too little.

What it does

• Live list. An agent shows up the moment it spawns, drawn like Claude Code’s own rows: type and description, its current tool call, how many more tool uses, then running time, model and tokens in and out. An agent’s own subagents sit indented under it.
• Cost per agent. Input tokens with the cached share, output, context size and request count, taken from the API’s usage figures for each response, plus the model and thinking effort its requests actually used. A response cut off at max_tokens shows in red.
• Conversations. Click an agent and the pane widens to show its brief, its replies as Markdown and its tool calls with their results, with runs of calls folded into one line. Scrolled back, the view holds still while the agent writes on.
• Timeline and receipt. The latest batch of agents on one time axis, a bar each, so you see what ran side by side and what took longest. When the batch ends, one line sums it up: wall time, total agent time and how parallel that was, tool uses and tokens.
• Stop. ■ Stop, then Confirm stop, ends a running agent through Claude Code’s own TaskStop.
• Other loops. Model loops no listed agent claims, such as a workflow’s agents or compaction and memory forks, are counted under the list.
• Opens and folds itself. The pane opens when an agent starts and folds 10 seconds after the last one ends, to a tab above the prompt. While it is folded or closed, a status line under the prompt counts the running agents, and a toast reports each one that finishes.

The same author’s statuspane does the session-wide view: model, context, rate limits and cost in a card above the prompt.

Install

Needs Claude Code 2.1.287 or later. The side pane docks only in the fullscreen layout (/tui fullscreen); on the main screen agentpane draws a summary of up to eight rows above the prompt instead.

claude plugin marketplace add xuanji86/claude-agentpane
claude plugin install agentpane@claude-agentpane

The pane opens by itself when an agent starts, as long as the terminal is at least 144 columns wide (110 once you have opened it yourself). /agentpane opens or closes it at any width. In a conversation, ▲ ▼ or k j step through it and ← or b goes back; those keys work while the pane holds the keyboard, which it takes when you open a conversation over an empty prompt. Esc gives the keyboard back.

In /config you can turn off the automatic opening, the toasts, the status line and the animation, set foldAfter (seconds, 0 keeps the pane open) and keepFinished (finished agents listed, 8 by default, up to 30).

How it works

agentpane learns of each agent through agent.spawn and polls the engine’s agent list every second, since no agent-status event reaches a mod. It hooks tool.call to read each agent’s tool name and a short argument (a Bash command, a file name, a search pattern, a URL) and turn.step for each response’s token usage and stop reason. It reads an agent’s transcript only while you have that conversation open, and strips control, bidi and zero-width characters from it before drawing.

Every hook passes its event on unchanged. It makes no network requests, runs no processes, reads and writes no files, stores nothing across sessions and calls no model. Its only action is Stop, which calls TaskStop after your second press.

Limitations

• Some figures are inferred, not measured. A start time is when the pane first saw the agent. Tokens count from when agentpane loaded, so an agent that finished before then shows none, and tool counts cover only the agents it watched.
• Workflow agents have no names the engine hands to mods. They count under other loops, where they cannot be told apart from compaction forks.
• A mod cannot switch Claude Code’s main view to an agent; the footer’s agent list does that (↓, then Enter). agentpane marks the agent the main view shows with ◂ main view.
• In the Claude Desktop Code tab, which has its own agents view, the code opens nothing by itself and shows no toasts or status line.