Claude Code Agents Status Bar

How many Claude Code agents are running right now? This extension answers that with one number in your status bar, and one click gives you the whole roster.

Unofficial extension. Not affiliated with, endorsed by, or supported by Anthropic. "Claude" is a trademark of Anthropic, PBC.

The problem

When Claude Code fans out, whether through a Task call, a Workflow spawning a dozen agents, or a background session left running in another window, that work becomes invisible. The official extension shows you the conversation you are looking at, not the fleet behind it. You end up wondering whether anything is still running, or whether you are burning tokens on agents you forgot about.

What you get

A single status bar item, always visible, spinning while work is in flight:

$(sync~spin) 3 agents

Click it and you get every one of them, grouped by the Claude session that spawned them:

Session [#1](https://github.com/LetermeFlorent/claude-subagents/issues/1)  a3f19c02
⟳ cavecrew-investigator  @claude-opus-5  e:high        42s
  [#1](https://github.com/LetermeFlorent/claude-subagents/issues/1) [hive] locate the worktree spawn path

⟳ general-purpose  @claude-sonnet-5                   3m18s  silencieux 8s
  [#1](https://github.com/LetermeFlorent/claude-subagents/issues/1) [hive] audit the fallback path

Session [#2](https://github.com/LetermeFlorent/claude-subagents/issues/2)  7b04ee51
⟳ bg  @claude-opus-5  e:medium                       17m04s
  [#2](https://github.com/LetermeFlorent/claude-subagents/issues/2) [resume-mail] watching CI on the docker branch

Agent type, model, reasoning effort, elapsed time, project and the description the agent was spawned with, all searchable since the quick pick matches on every field. Durations count up live while the list is open. Pressing Enter on an agent opens its transcript.

Side panel

A quick pick closes the moment you click anywhere else, which is the wrong shape for something you want to keep an eye on. Set claudeSubagents.listStyle to panel and the list moves into its own view in the activity bar, where it stays put:

CLAUDE AGENTS                            3

SESSION [#1](https://github.com/LetermeFlorent/claude-subagents/issues/1)                               2
│ cavecrew-investigator             42s
│ locate the worktree spawn path
│ claude-opus-5 · high · hive

│ general-purpose                  3m18s
│ audit the CurseForge fallback
│ claude-sonnet-5 · hive · silencieux 8s

SESSION [#2](https://github.com/LetermeFlorent/claude-subagents/issues/2)                               1
│ bg                              17m04s
│ watching CI on the docker branch
│ claude-opus-5 · medium · resume-mail

Each agent gets three lines rather than one: what it is and how long it has been running, what it was actually asked to do, then its model, effort and project. The rail down the left carries the only thing you really scan this panel for, whether the agent is still moving: it stays green while output keeps coming, turns yellow after five seconds of silence and orange past two minutes, and blue marks a remote agent. Measured values are set in the editor font, prose in the interface font, so numbers line up and read apart from descriptions at a glance.

Sessions only get headers when there is more than one, the count rides as a badge on the activity bar icon, and clicking an agent opens its transcript. Durations tick every second from absolute timestamps computed in the view itself, so nothing is re-read from disk to keep them moving. With the panel on, clicking the status bar counter reveals it instead of opening the quick pick.

Telling three sessions apart

Claude Code stores subagents under the session that spawned them, at ~/.claude/projects/<project>/<session>/subagents/. Running three Claude sessions in one project therefore produces three separate trees, and each agent belongs to exactly one of them. The list groups agents under their parent session and gives each one a short number, so you can tell which of your windows is doing what.

Features

Everything gets counted, not just the obvious: Task and Agent subagents, agents spawned inside a Workflow, and background sessions started with claude agents or claude --bg.

Thinking time does not kill an agent. Workflow agents are tracked through the run journal rather than file timestamps, so one that reasons for four minutes without writing anything still counts as alive.

The list is scoped to your workspace by default, so a busy machine does not pollute the window you are working in, and one setting shows everything. Remote machines work too, over plain SSH, which lets you watch the agents running on a VPS from your laptop.

No dependencies and no network access. The extension reads files Claude Code already writes, and transcript headers are cached between scans, so a steady state costs no disk reads at all.

How it works

Claude Code records every agent on disk. The extension reads those records every few seconds and decides what is still alive:

The journal rule is the important one. A thinking agent writes nothing for minutes, so the naive "was the transcript touched recently?" check reports most of a workflow as dead.

CLAUDE_CONFIG_DIR and CLAUDE_PROJECTS_DIR are honoured if you have moved your Claude directory.

Watching remote machines

Add SSH aliases from your ~/.ssh/config:

"claudeSubagents.remoteHosts": ["growthfit", "syllabis-vps"]

Each host is polled every 15 s with ssh -o BatchMode=yes (6 s timeout), and its agents appear in the list prefixed with the host name. Requires ssh on your PATH and a key without a passphrase, since batch mode never prompts: it fails silently and the host is skipped.

Settings

Commands

Troubleshooting

If it always says 0 agent, remember that only agents whose working directory is inside the current workspace are counted by default. Turn on claudeSubagents.showAllProjects to check whether that filter is what hides them.

If remote agents never appear, test the exact command the extension uses: ssh -o BatchMode=yes <alias> true. Anything that prompts or fails leaves the extension with nothing.

Two items in the status bar over Remote-SSH should not happen, since the extension declares "extensionKind": ["ui"] and only ever runs in the local extension host. If you installed an older build server-side, uninstall it there.

Privacy

No telemetry, no analytics and no network access, apart from the ssh calls to hosts you explicitly list. Everything else is local file reads.

License

MIT