一群 Claude Code 同时开工，
谁做了什么，大家都知道。 Many Claude Code sessions.
One shared memory.

实际用法How it's used

从 Jira 里现成的 ticket，到一组在干活的 sessionFrom tickets already in Jira to a set of working sessions

Butler 不替你规划，也不替你拆需求。它做的是把「一个 ticket 一个 Claude Code session」这件事管起来：开 session、分组、看状态、记下和传递上下文。下面是一个 epic 从规划到开工的真实流程。 Butler doesn't plan for you or split requirements for you. It runs the "one Claude Code session per ticket" way of working: opening sessions, grouping them, showing their state, and keeping and passing on context. This is how an epic actually goes from plan to work.

1. 1

   ticket 早就在 Jira 里The tickets are already in Jira

   产品已经把需求写成了 ticket。Butler 不拆 ticket，也不连 Jira：读 Jira 的是你自己的 Claude Code，通过它自己的 Jira MCP。Product has already written the tickets. Butler doesn't split them and doesn't connect to Jira: your own Claude Code reads Jira, through its own Jira MCP server.

2. 2

   在 epic session 里理成几个 featureSort them into features in an epic session

   在 epic 页打开一个 epic session，和 Claude 一起把这些 ticket 归成几个 feature，理清先后，以及前后端怎么配合。定下来的 feature 由 epic session 建出来，附上一份规划：做什么、为什么、已经定了什么。Open an epic session on the epic page and work out with Claude which tickets form which feature, in what order, and how frontend and backend fit together. When a feature is settled, the epic session creates it with a short plan: what to build, why, and what is already decided.

   create_feature

3. 3

   打开 feature 的总控，开工Open the feature's coordinator and start

   点进 feature，打开它的总控。总控第一次打开时会收到这份规划，以及同一个 epic 里其他 feature 的进展。它建好 ticket，再给每个 ticket 开一个 session，把任务发过去。一个 feature 最多同时跑 8 个终端。Click into the feature and open its coordinator. The first time, it receives the plan plus the progress of the epic's other features. It creates the tickets, then opens a session in each one and hands it a task. A feature runs at most 8 terminals at once.

   create_ticket start_session

4. 4

   你只处理「需要你」的You handle what needs you

   「现在」页把每个 session 按 epic / feature / ticket 排好，等你回答、等你确认、或者卡住了的，会置顶。每晚 Butler 用你自己登录的 claude 给有进展的 ticket 写一份交接（handoff），第二天的 session 一 recall 就接得上。The Now page lists every session by epic, feature and ticket, and pins the ones waiting for an answer, a confirmation, or stuck. Every night Butler asks your own logged-in claude to write a handoff for each ticket that moved, so the next day's session picks up from it with one recall.

三层记忆Layered memory

「大家都知道」是怎么做到的How "everyone knows" actually works

Butler 内置了本地的 StateCore 记忆，给它打开的每个 session 都接上几份记忆。session 用 remember 记下，用 recall 读到。总控不会往别的 session 的终端里打字，大家知道彼此做了什么，靠的就是这几份共用的记忆，一层套一层。 Butler embeds a local StateCore memory and connects every session it opens to a few memories. A session writes with remember and reads with recall. A coordinator never types into another session's terminal. What keeps everyone in sync is these shared memories, nested one inside the next.

一个具体的例子A concrete example

后端 session 改完接口就记下一句，同一个 feature 里的前端 session 下次 recall 时读到，照着改。中间没有人转述，总控也不用去敲前端的终端。The backend session notes the API change once it lands; the frontend session in the same feature reads it on its next recall and follows it. Nobody relays it, and no coordinator has to type it into the frontend's terminal.

「现在」页The Now page

打开 Butler，先看到的是需要你的Open Butler and see what needs you first

首页「现在」列出 Butler 开的每一个 session，按 epic、feature、ticket 分组。每一行写着它此刻在做什么：写代码（哪个文件）、跑测试、等 CI、思考中，后面是它的 PR 和 CI，以及这个状态持续了多久。The home page, Now, lists every session Butler opened, grouped by epic, feature and ticket. Each row says what the session is doing right now (writing code and in which file, running tests, watching CI, thinking), followed by its PR and CI and how long it has been at it.

等你的 session 置顶在「需要你」里，原因只有三种：它问了你一个问题、它停在权限确认上、或者它卡住了。侧栏的「现在」后面会显示有几个。Sessions waiting on you are pinned at the top under Needs you, for one of three reasons: they asked you a question, they stopped on a permission prompt, or they are stuck. The sidebar shows how many there are.

点一行就打开它的终端，接手、回答或者批准都在那里。没在干活的 session 会折起来，页面底部有一行近 7 天的 token 用量。Click a row to open its terminal, where you take over, answer or approve. Idle sessions fold away, and the bottom of the page has one line of token usage for the last 7 days.

演示数据Demo data

Token 用量Token usage

一行字，看出 token 花在了哪儿One line that shows where the tokens went

一群 session 同时跑，最容易问的就是「这周用了多少，都用在哪了」。「现在」页最底下有一行近 7 天的用量，点开能看到按类型和按模型的拆分。 With many sessions running at once, the obvious question is how much this week used, and on what. The bottom of the Now page has one line for the last 7 days; click it for the split by kind and by model.

本地优先Local-first

它是你电脑上的一个应用，不是一项服务An app on your computer, not a service

Butler 不在你和 Claude Code 之间加一层服务器。它管的是你自己的 claude，用的是你自己的登录。 Butler puts no server between you and Claude Code. It manages your own claude, signed in as you.

安装Install

下载、放行、跑一遍向导Download, allow it, run the wizard

安装包在 GitHub 的 Releases 页上。目前的构建没有代码签名，所以浏览器和系统都会提醒你，下面写了怎么放行。 The installers are on the GitHub Releases page. The builds are not code-signed yet, so your browser and OS will warn you; here is how to allow them.

xattr -dr com.apple.quarantine /Applications/Butler.app

如果你也是一个人开着一堆 Claude Code，可以试试看。If you also keep a pile of Claude Code sessions open, give it a try.