agentpane

See your subagents at work, beside the conversation.

Every subagent a Claude Code session runs, what it is doing right now, what it costs in tokens, and its whole conversation one click away.

English · 中文

Why

When Claude fans work out to subagents, the terminal shows one line per agent and a footer list. agentpane is a
mod that keeps them in a side pane, drawn the way Claude
Code draws its own rows; click one and Claude Code's main view shows its conversation.

Features


Live list	`⏺ Type(description)`, the tool call it runs now, `+N more tool uses`, and `✶ Running… (1m 12s · Opus 5.5 · 412k in · 1.9k out)`, or `✻ Waiting…` while it waits on its own background work; listed the moment it spawns
Nested agents	An agent's own subagents sit indented under it
Tokens, model and effort	Input (cached share shown), output, context size and requests per agent, from the API's own usage figures; the model and thinking effort each agent's requests actually use; `max_tokens` in red when a response was cut off
Other loops	Model loops no listed agent claims (a workflow's agents, compaction and memory forks), counted under the list
Timeline	The latest batch of agents on one time axis, a bar each: what ran side by side, and what took longest
Batch receipt	Once a batch ends: `✓ 4 agents in 31s · 1m 18s of agent time (2.5× in parallel) · 21 tool uses · 190k tokens`
Status line	`✻ 2 of 4 agents running` under the prompt while the pane is folded or closed
Follows the main view	The agent you open from Claude Code's footer list is marked `◂ main view`
Conversations	Click an agent: the main view switches to its transcript, and `◂ Main` in the pane's header goes back. Where the main view can't switch (Claude Code before 2.1.300), it opens in the pane: its brief, replies in Markdown, tool calls with their results, runs of calls folded into one line
Stop	`■ Stop` on a running agent's row, then `Confirm stop`, ends it through Claude Code's TaskStop
Notices	A toast when an agent finishes, one for several at once
Opens and folds itself	Opens when an agent starts, folds 10 s after the last one ends, to a tab above the prompt
Theme colours	Claude Code's own theme keys: follows dark, light and colourblind themes
Light on redraws	The spinner, shimmer, clocks and timeline run on the terminal's own frame clock; the pane redraws only when something changes
Main screen too	Without fullscreen, a summary of up to eight rows above the prompt instead of a side pane

Install

Needs Claude Code 2.1.287 or later, the release that brought mods (function hooks); tested on 2.1.300. The side
pane needs the fullscreen layout (/tui fullscreen); on the main screen agentpane is a summary above the prompt. Mods
are early access: their API may change between releases, and a release that breaks the pane gets a fix here.

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

From a terminal instead

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

Use

A conversation in the pane — before Claude Code 2.1.300

One agent's conversation in the widened pane

Where the main view can't switch, clicking an agent opens its conversation here, and the pane widens to read it. ▲ ▼ step through it (k j), ← goes back (b); scrolled back, the view holds
still while the agent writes on. Click a folded run of tool calls to open it.

Folded — the ▸ handle on the pane's left edge

The pane hides and a tab above the prompt keeps the counts. Click it to bring the pane back; a new agent brings it back by itself.

/agentpane opens or closes the pane.
3 done in the header hides the finished agents (their running children stay in sight).
The pane opens by itself where Claude Code seats a pane nobody asked for: at least 144 columns (110 once you have
opened it yourself). Below that, /agentpane opens it.
The keys b k j work while the pane holds the keyboard: it takes it when you open a conversation (over an empty
prompt), Esc gives it back.

Configure

In /config, or under pluginConfigs["agentpane"].options in settings.json:

Option	Default	Meaning
`autoOpen`	`true`	open the pane when an agent starts
`foldAfter`	`10`	seconds after the last agent finishes before the pane folds; `0` keeps it open
`motion`	`true`	animate the spinner and shimmer; off, they stand still and the clocks still run
`toasts`	`true`	a toast when agents finish
`keepFinished`	`8`	finished agents the list keeps (1–30)
`statusLine`	`true`	the running agents under the prompt while the pane is not on screen

What it can reach

Its footprint, as the awesome-claude-code-mods scan reads it from claude plugin validate:

agentpane watches, and acts only when you click an agent or press Stop. Every hook passes its event on unchanged.

It sees	Through
the session's agents and their status	`$.agent.list()`, `agent.spawn`
each agent's tool calls (name and a short argument)	`tool.call`
each agent's token usage and stop reason	`turn.step`
an agent's transcript, while you have it open	`$.session.messages({ agentId })`
which agent the main view shows	the pane's `view`
the main view, switched to an agent you click (`◂ Main` back)	`$.ui.view`

What it draws outside its pane: the tab above the prompt, toasts, and one status line ($.ui.status).

It makes no network requests, runs no processes, reads and writes no files, stores nothing across sessions and calls
no model. Its actions are yours: switching the main view on your click, and Stop: Claude Code's own TaskStop, on your second press, naming the agent in a cleaned,
shortened line. Text from transcripts is stripped of control, bidi and zero-width characters and cut to length before it
is drawn. claude plugin validate . prints exactly what it hooks and calls.

What is inferred, not measured

Start times are when the pane first saw the agent: at its spawn, or, for one already running when agentpane
loaded, the next poll.
Tokens count from when agentpane loaded; an agent that finished before then shows none.
Other loops cannot tell a workflow's agent from a compaction fork: both are model loops no listed agent claims.
Tool counts are the calls the pane saw, and show only for agents it watched.
A batch is the agents that ran with no quiet minute between them; the timeline and the receipt cover the latest.
Agent time is the sum of their run times, so 2.5× in parallel is that sum over the batch's wall time.

Limits

The main view switches only on your click, and only from Claude Code 2.1.300 ($.ui.view). Before that, and in
Claude Desktop until its own copy of Claude Code catches up, the conversation opens in the pane instead.
Workflow agents have no names to list: the engine does not hand them to mods. They count under other loops.

Troubleshooting

The pane doesn't appear. Check claude --version is 2.1.287 or later, run /reload-plugins, then /agentpane.
Below 144 columns Claude Code won't seat a pane nobody asked for; /agentpane opens it at any width. Look in the
transcript for a dim line starting agentpane:: it names a hook that failed or why the pane was refused. Please
open an issue with it.

b k j do nothing. The pane needs the keyboard: click in it, or open a conversation over an empty prompt.

Too much motion. Set motion to false in /config.

Develop

git clone https://github.com/xuanji86/claude-agentpane
cd claude-agentpane
claude plugin validate .
claude plugin test .
claude --plugin-dir .        # or add the folder to CLAUDE_CODE_PLUGIN_DIRS

A session that loaded the mod from its folder reloads it each time you save a file. The tests draw the pane on every
surface (terminal, desktop, VS Code, mobile) at 40–120 columns. python3 assets/make_previews.py redraws the README
pictures. Changes are listed in CHANGELOG.md. Issues and pull requests are welcome.

claude-agentpane