Features
- Added Traditional Chinese (繁體中文) support with Taiwan wording, automatically selected for zh-TW, zh-HK, and zh-Hant systems.
- Added “Route by what it asks for” for intent-based routing, with Jev as the classifier for groups whose effort Jev picks.
- Added an official
magpie-quotaskill to the skills market, so agents can check remaining allowance and wait for quota. - Added GLM Coding Plan reset cards for bigmodel.cn and Z.ai, showing five-hour and weekly resets counted the same way as the vendor.
Bug Fixes
- Fixed usage windows that had passed their reset not renewing consistently; accounts now return automatically once a reading finds the renewed window.
- Fixed Claude usage resets without a year being read as a year away, which could bench an account for a year.
- Fixed clicks made while a page was redrawing being lost, including settings presses and menus opening from the redrawn button.
- Fixed a plugin host lifecycle issue after failed writes so a freed call doesn’t reference an already-ended host.
- Fixed Devin’s Claude models returning 502 to Codex desktop’s automation_update when tool parameters use root
anyOf/oneOf/allOf; they now send as a plain object. - Fixed skill installation so already-installed skills are skipped, the rest install, and the toast names both installed and skipped skills.
- Fixed the Library page going blank after Check for updates for skills or servers switched off for every agent; they are now saved as
"agents": []instead of null. - Fixed Magpie web’s login shell PATH detection suspending the terminal; it now runs in its own session.
- Fixed OpenCode Go and Zen providers saved with only a chat URL: requests now use the correct API endpoint, and errors say which URL was asked.
- Fixed subscription window layout causing a ResizeObserver loop error in WebKit.
- Fixed the Requests table account column cutting off “Subscription” in narrow windows.
Improvements
- Magpie now tracks context window details for every request, session, and agent.
- Volcengine plans now use their preset models instead of fetching them from the
/modelsendpoint. - Linux CI now times out stuck GTK/WebKitGTK installation after 15 minutes instead of holding builds for hours.
Install
Download from usemagpie.ai, or curl -fsSL https://usemagpie.ai/install.sh | sh. An installed magpie updates itself.
- macOS (Apple Silicon):
magpie-darwin-arm64.dmg - macOS (Intel):
magpie-darwin-amd64.dmg - Windows:
magpie-windows-amd64.exe(ARM:magpie-windows-arm64.exe) - Linux:
magpie-linux-amd64/magpie-linux-arm64(needs GTK 3 and WebKitGTK 4.1) - Terminal only:
magpie-cli-<os>-<arch>
功能
- 添加了繁体中文(繁體中文)支持,采用台湾用语,并会针对 zh-TW、zh-HK 和 zh-Hant 系统自动选择。
- 添加了“Route by what it asks for”用于基于意图的路由,并以 Jev 作为分类器,用于 Jev 选择 effort 的分组。
- 在技能市场添加了官方
magpie-quota技能,以便 agent 可以检查剩余额度并等待配额。 - 为 bigmodel.cn 和 Z.ai 添加了 GLM Coding Plan 重置卡片,显示以与供应商相同方式计算的五小时和每周重置。
错误修复
- 修复了已过重置时间的使用窗口未能一致续期的问题;现在,一旦读取发现已续期窗口,账户会自动恢复。
- 修复了 Claude 使用量重置在没有年份时被读取为距今一年后,可能导致账户被停用一年的问题。
- 修复了页面重绘期间进行的点击丢失的问题,包括设置点击和从被重绘按钮打开的菜单。
- 修复了写入失败后插件宿主生命周期的问题,使已释放的调用不再引用已结束的宿主。
- 修复了 Devin 的 Claude 模型在工具参数使用根级
anyOf/oneOf/allOf时向 Codex 桌面的 automation_update 返回 502 的问题;现在它们会以普通对象发送。 - 修复了技能安装逻辑:已安装的技能会被跳过,其余技能正常安装,提示中会列出已安装和被跳过的技能名称。
- 修复了当所有 agent 都关闭了技能或服务器的“Check for updates”后 Library 页面空白的问题;现在它们会保存为
"agents": []而不是 null。 - 修复了 Magpie web 的登录 shell PATH 检测导致终端挂起的问题;它现在会在自己的会话中运行。
- 修复了仅保存聊天 URL 的 OpenCode Go 和 Zen 提供商的问题:请求现在会使用正确的 API 端点,错误信息会说明请求的是哪个 URL。
- 修复了订阅窗口布局在 WebKit 中导致 ResizeObserver 循环错误的问题。
- 修复了 Requests 表格中账户列在窄窗口中截断“Subscription”的问题。
改进
- Magpie 现在会跟踪每个请求、会话和 agent 的上下文窗口详细信息。
- Volcengine 套餐现在使用其预设模型,而不是从
/models端点获取。 - Linux CI 现在会在 15 分钟后使卡住的 GTK/WebKitGTK 安装超时,而不再让构建阻塞数小时。
安装
从 usemagpie.ai 下载,或运行 curl -fsSL https://usemagpie.ai/install.sh | sh。已安装的 magpie 会自动更新。