Features
- Added an IP/network allow list for the gateway so you can control which addresses can connect. Configure it with
MAGPIE_ALLOW_IPSor in Settings next to Share on local network. - Added Hooks as a new Library category. Hooks are managed once and written into the selected agents' hook files.
- Added a Strict stay option that keeps a conversation on the same account or key that answered it, even after hitting its cap, and passes errors on instead of moving the conversation.
- Usage now includes CodeBuddy Code's own calls by reading its session files, similar to WorkBuddy.
- When web search falls back, the Routing row now shows which search service failed or found nothing, and why.
Bug Fixes
- Fixed Windows taskbar showing a blank icon instead of the app icon.
- Codex subagents now see all available models, not just the first five.
- Custom context windows set on a provider or routing group are now used by Codex and Claude Code.
- Fixed plugin host hangs on Windows and excessive memory use on Linux when a plugin prints very long output.
- Copilot Claude models with dotted IDs now appear to Claude Desktop/Code with dashed IDs.
- Providers set to "Select none" no longer serve unlisted models.
- Volcengine and Tencent keys are kept through sync, restore, copy, and stale editor sessions, and plan cards now tell you what key is needed when one is missing.
- Qoder queue responses now show as retryable 503 responses instead of "sign in again".
- WorkBuddy models now search with WorkBuddy's own search service, so Claude Code WebSearch gets results.
- Image recognition and web search now skip models that are out of allowance.
- Requests with no reply limit now use the model's own limit on Anthropic's API instead of a fixed limit.
- Copilot's Claude models no longer fail when the provider refuses auto safeguards; requests are retried without them.
- Kimi conversations started by another model no longer fail with 400 errors on later turns.
- Out-of-allowance keys are no longer retried on every request, and duplicated keys are only asked once.
- Oversized MCP requests are now refused with 413 instead of being partially processed.
- Codex proxy detection now uses login shell and macOS network proxy settings, avoiding false proxy warnings.
- Invalid 512-token window listings no longer cause agents to receive the wrong model window.
- Accounts out of credits can still use free models included with their subscription.
- OpenCode Zen now clearly explains that an empty additional key field adds a public key.
- Codex threads started through CC Switch's OpenAI mirror are no longer refused when reopened with a Magpie model.
- Fixed default model selection for OMP when an enabled model list exists.
- When settings cannot be read, the gateway now keeps the previous allow list or allows only this computer, never every address.
- Windows plugins added from a folder are now found by their short name in CLI commands.
- Effort settings now last across model switches in Codex, Claude Code, and Grok Build.
Improvements
- Hermes Agent now uses the correct local app data folder on Windows.
- The Library Backups button is now at the top of every Library tab for easier access.
- Settings > General now uses simpler labels and a single Language menu.
- Command path lookups are cached on Windows, reducing repeated filesystem access.
- Models added to Codex subagents are named neutrally.
- Providers can pass Codex requests through unchanged when configured to do so, for relays that expect Codex's own requests.
- Refusal messages now keep Anthropic's stop reason details in Routing and agent errors.
- Disabling subscription accounts now asks for confirmation and lists the subscriptions that would stop working.
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>
Features
- 为网关添加了 IP/网络允许列表，以便控制哪些地址可以连接。可通过
MAGPIE_ALLOW_IPS或在“设置”中“在本地网络上共享”旁边进行配置。 - 将 Hooks 添加为库中的一个新类别。Hooks 只需管理一次，并会写入所选代理的 Hook 文件。
- 添加了 Strict stay（严格停留）选项，对话会保持在应答它的同一账户或密钥上，即使在达到上限后也会继续传递错误，而不是转移对话。
- 用量现在会通过读取 CodeBuddy Code 的会话文件，将其自身的调用也计入，类似 WorkBuddy。
- 当网页搜索发生回退时，Routing 行现在会显示哪个搜索服务失败或未找到结果，以及原因。
Bug Fixes
- 修复了 Windows 任务栏显示空白图标而非应用图标的问题。
- Codex 子代理现在可以看到所有可用模型，而不仅仅是前五个。
- 在提供商或路由组上设置的自定义上下文窗口现在会被 Codex 和 Claude Code 使用。
- 修复了当插件输出非常长的内容时，Windows 上插件宿主挂起以及 Linux 上内存占用过高的问题。
- ID 中带点号的 Copilot Claude 模型现在会以带短横线的 ID 显示给 Claude Desktop/Code。
- 设置为“Select none”的提供商不再提供未列出的模型。
- Volcengine 和 Tencent 的密钥会在同步、恢复、复制和过期的编辑器会话中保留，方案卡片现在会在缺少密钥时告诉你需要哪个密钥。
- Qoder 队列响应现在会显示为可重试的 503 响应，而不是“sign in again”。
- WorkBuddy 模型现在会使用 WorkBuddy 自己的搜索服务进行搜索，因此 Claude Code WebSearch 能获得结果。
- 图像识别和网页搜索现在会跳过额度已用尽的模型。
- 未设置回复上限的请求现在会使用模型在 Anthropic API 上的自身上限，而不是固定上限。
- Copilot 的 Claude 模型不再因提供商拒绝自动安全防护而失败；请求会在不带有这些防护的情况下重试。
- 由其他模型发起的 Kimi 对话不再在后续轮次中因 400 错误而失败。
- 额度已用尽的密钥不再在每次请求时都重试，重复密钥也只会询问一次。
- 过大的 MCP 请求现在会以 413 拒绝，而不是被部分处理。
- Codex 代理检测现在使用登录 shell 和 macOS 网络代理设置，避免出现错误的代理警告。
- 无效的 512 token 窗口列表不再导致代理收到错误的模型窗口。
- 额度已用尽的账户仍可使用其订阅中包含的免费模型。
- OpenCode Zen 现在清楚地说明，将附加密钥字段留空会添加一个公钥。
- 通过 CC Switch 的 OpenAI 镜像启动的 Codex 线程，在使用 Magpie 模型重新打开时不再被拒绝。
- 修复了存在已启用模型列表时 OMP 的默认模型选择问题。
- 当设置无法读取时，网关现在会保留之前的允许列表，或仅允许本机访问，绝不会允许所有地址。
- 从文件夹添加的 Windows 插件现在可以在 CLI 命令中通过其短名称找到。
- Effort 设置在 Codex、Claude Code 和 Grok Build 中切换模型后现在会保留。
Improvements
- Hermes Agent 现在在 Windows 上使用正确的本地应用数据文件夹。
- 库中的“备份”按钮现在位于每个库标签页的顶部，便于访问。
- 设置 > 通用现在使用更简单的标签和单一“语言”菜单。
- Windows 上的命令路径查找现在会被缓存，减少重复的文件系统访问。
- 添加到 Codex 子代理的模型现在会使用中性命名。
- 提供商可以配置为原样传递 Codex 请求，适用于期望接收 Codex 自身请求的中继。
- 拒绝消息现在会在 Routing 和代理错误中保留 Anthropic 的停止原因详情。
- 禁用订阅账户现在会要求确认，并列出将停止工作的订阅。
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。
The same release, and the feed magpie updates from: https://github.com/yetone/magpie-releases/releases/tag/v0.1.1214