hivemind

hivemind shows your subagents as a live tree in a sidebar. The main loop sits at the root, every subagent hangs under the loop that spawned it, and each node says what that agent is doing right now, what it has cost so far and how full its context is. /hivemind opens and closes it.

What it does

• Status and task. Each node has a status glyph (a spinner while it runs, ✓ done, ✖ failed, ⏸ waiting), the subagent type, its task description, ⇢bg for background agents and the elapsed time.
• Numbers. A second line carries model and effort, requests, tokens, an estimated spend and the context fill of the agent’s last request as a small bar.
• Current tool. A third line names the tool the agent runs now, with a short detail such as a file name or a command’s description and a timer, or the last tool it ran and how long it took.
• Header. It counts running, done and failed agents, totals their spend, tokens and requests, and draws a sparkline of how many ran at once.
• Folding. f hides finished branches; a finished agent stays visible while one of its children still runs.

In the Code tab of Claude Desktop the header becomes a row of tiles over an activity chart, and the tree is one drawing with connector lines that flow while a child runs, in the light and the dark theme.

Install

Needs Claude Code 2.1.290 or later.

claude plugin marketplace add ersinkoc/claude-mods
claude plugin install hivemind@kozmos

Run /reload-plugins or start a new session, then /hivemind. Until Claude starts a subagent the tree holds only main. The autoOpen option, off by default, opens the sidebar when a session starts.

The kozmos marketplace holds the 48 mods of the KOZMOS collection; claude plugin install kozmos@kozmos adds a /kozmos hub that lists them and opens each one with a click. One of them, orrery, draws the same subagents as planets above the prompt.

How it works

The module follows subagent spawns (agent.spawn), every model request (turn.step, for model, effort, tokens and context), turn ends and every tool call, from which it takes the tool name and a short detail to display. While anything runs it ticks once a second for the timers and reads Claude Code’s own agent roster every three seconds, which wins on status and parentage. It makes no network calls, writes no files and adds nothing to the prompt.

Limitations

• Spend per agent is an estimate from list prices per model. For the engine’s own session total the author points to KOZMOS Bridge.
• After a plugin reload the tree starts over. The roster brings back the agents Claude Code still knows, without their earlier tokens and spend.
• The author warns that the mod API may change between Claude Code releases.