--- title: "agent-crew" url: "https://claudemods.dev/zh-cn/builds/agent-crew-6ba7e2" source: "https://github.com/bekir1184/agent-crew" source_type: "github" repo: "https://github.com/bekir1184/agent-crew" author: "bekir1184" author_handle: "bekir1184" stars: 6 category: "仪表盘与 HUD" categories: ["hud"] tags: ["cost-tracking", "hooks", "monitoring", "pixel-art", "statusline", "subagents"] permission_level: 2 scan_complete: true install: "claude plugin marketplace add bekir1184/agent-crew\nclaude plugin install agent-crew@agent-crew" added: "2026-10-09T20:08:19.970Z" updated: "2026-10-09T15:55:51Z" locale: "zh-cn" content_locale: "zh-cn" --- # agent-crew 在提示符上方以像素艺术风格的 Claude 团队成员展示你的子代理：每个子代理在做什么、进展如何以及使用了多少 token。 ## Description Agent Crew 是一个 Claude Code 插件，它会在提示符上方将你的子代理渲染成一个小小的像素艺术团队。它通过 `/plugin install agent-crew --marketplace bekir1184/agent-crew` 安装，并在子代理运行时自动工作。主要功能： - 一个带边框的 AGENT 行，显示 Claude 的实时活动（搜索、阅读、思考、写作、运行测试）以及一行包含 token 和工具计数的回合摘要。 - 每个子代理一行，标注其运行的模型，辅助代理嵌套在其父代理下。繁忙时行会折叠。 - 针对每个正在使用的工具显示动画小生物状态，不同的思考动画，以及当代理等待许可时显示的黄色“需要批准”气泡。 - 每行可展开的详细面板：完整任务、待办事项列表、涉及的文件、token 细分以及运行摘要。 - 根据过去运行时长估算的剩余时间，无需模型调用，无 token 成本。 - 诚实的 token 核算：只显示代理自身的读/写 token，缓存和重读 token 保留在单独详情行中。 - 通过设置可选的估算成本显示。命令：`/crew`（切换可见性）、`/crew demo`（30 秒演示）、`/crew clear`（重置舞台）。设置位于 `/plugin configure agent-crew@agent-crew`。在终端中，行显示为彩色圆点，可通过 `ctrl+x tab` 导航。开发：使用 `claude --plugin-dir .` 运行，使用 `claude plugin validate .` 验证，使用 `claude plugin test .` 测试。动画文档图片通过 `node scripts/docs.mjs` 重绘。MIT 许可证，需要 Claude Code >= 2.1.293。 ## Install ``` claude plugin marketplace add bekir1184/agent-crew claude plugin install agent-crew@agent-crew ``` ## Original README

Agent Crew

Your subagents, as a tiny pixel crew above the prompt.
See who is doing what, how far along they are, and how long is left.

Agent Crew in the Claude desktop app: Claude starts alone on the AGENT line, then a Plan, an Explore and a General agent arrive one by one with a helper; they search, read, think and ask for approval, their bars fill up, and each raises a checkered flag when done

## Install ``` /plugin install agent-crew --marketplace bekir1184/agent-crew ``` That's it. The crew shows up by itself whenever Claude is at work. Want a look right now? Type `/crew demo`. **Costs nothing.** Agent Crew never calls a model and never touches your prompts: it only watches. ## Claude, live While Claude works, the framed line shows what it's doing this very second. When the turn ends it shrinks to one quiet line with the totals. Press the arrow to see what happened. The AGENT line: searching, reading, thinking, writing, running tests, then a thin line: done, last turn 0:37, 5.6k tokens, 9 tools

The AGENT line: searching, reading, thinking, writing, running tests, then a thin line: done, last turn 0:37, 5.6k tokens, 9 tools

## The crew When Claude hands work to subagents, they line up below, one row each. Each row names the model it runs on, so you can tell Opus from Haiku at a glance. Helpers started by a subagent tuck in under their parent. Too busy? The arrow on the title folds them away. Each critter holds the tool it's using: Searching with a magnifier, reading a page, writing in a code window, running a command on a laptop, browsing a globe, getting ready in a hard hat, done with a checkered flag, failed with dizzy eyes, cancelled and asleep

## Thinking... Thinking is where agents spend most of their time, so the critters keep it lively: an hourglass, a thought cloud with a question mark, a lightbulb that lights up. Each one at its own pace. Three critters thinking: one turns an hourglass, one fills a thought cloud with a question mark, one lights a bulb and hops

Three critters thinking: one turns an hourglass, one fills a thought cloud with a question mark, one lights a bulb and hops

## It needs you When an agent stops to ask for permission, it says so: a yellow question bubble, and the title tells you how many are waiting. A row turns yellow with Needs approval: Bash rm -r build, then carries on once approved

A row turns yellow with Needs approval: Bash rm -r build, then carries on once approved

## The details Every row has an arrow. It opens everything the row had no room for: the full task, its to-do list, the files it touched, and its tokens. A row opens into a panel: task, current step, checklist, changed files, token breakdown and run summary

A row opens into a panel: task, current step, checklist, changed files, token breakdown and run summary

## How long is left? No model is asked and no tokens are spent: Agent Crew simply remembers how long this kind of agent took before. Past run times, the middle one picked, a live countdown, and a re-estimate when the run goes long

Past run times, the middle one picked, a live countdown, and a re-estimate when the run goes long

## Honest tokens A row shows what the agent's own work took: what it read in and what it wrote. Nothing else. Claude also keeps the whole conversation in a cache and re-reads it on every request. And after a break the cache expires, so the next request writes the whole conversation into it again, which on a long conversation can be hundreds of thousands of tokens at once. Neither is the agent's work, so neither inflates the row: both sit in the details panel, on their own line. Turn on **Show estimated cost** in the settings and each row also gets a ≈ dollar figure. ## Commands | | | | --- | --- | | `/crew` | Hide or show the crew | | `/crew demo` | A 30-second show: Claude starts alone, the crew arrives, thinks, asks you, and ends with all flags up | | `/crew clear` | Clear the stage | **Settings** live in `/plugin configure agent-crew@agent-crew`: show Claude's own line (on), show estimated cost (off). **In the terminal** there's no pixel art, just a colored dot per row. Press `ctrl+x tab` to reach the crew, then the letter on an arrow: `h` for the title, `a`, `b`, `c` for the rows. ## Development Run it from a clone, then check and test it (Claude Code 2.1.293 or newer): ```bash claude --plugin-dir . claude plugin validate . claude plugin test . ``` The animated images in `docs/` and `site/` are drawn from the mod's own pixel art: `node scripts/docs.mjs` (Node 25+) redraws them. ---