Changes
- gateway tests: TestClaudeRunEndsAfterStructuredOutput waits for the bridge to let its runs go, not only their processes
- provider: moved WorkBuddy users get plugin 0.1.20, so Claude Code's tool results reach WorkBuddy AI's gemini-3.8-flash (magpie-community/plugins#113)
- gui: the tray's quota line is taller and shades its gap to the even burn (#2079)
- access, gui, cli: a gateway key's limit can run every N hours, a 5-hour cycle among them (#2102)
- agent: Claude Code's fable tier is left unset on another vendor's model too (#2090)
- provider: a window set for all of a provider's models is no larger than a model's own (#2112)
- gateway: Gemini's tags stay out of the answer when only the opening chunk is marked (#2098)
- gateway: Gemini 4 and later are told a thinking level, at the level asked (#2083)
- gui: a gateway key's account menu says a provider none of whose accounts is picked is used, and how to keep the key off it (#2077, KKKKKKKEM)
- gateway: the key's own limit on a remote magpie still rests the whole remote, and #2077's case is tested across two gateways (#2077, KKKKKKKEM)
- fix(catalog): honor models.dev reasoning toggles (#1294)
- gateway tests: a Chat client's reply translated from Anthropic's Messages streams as the vendor sends it — the first chunk when its first event comes, each word as it is written, never the whole reply held to the stream's end — and now a test holds it to that. A reporter's OpenAI-format client on Z.ai's GLM saw the opposite for a while (every event of a reply in one burst once the stream ended, while the same provider's Messages replies streamed fine), and it measured out to be the vendor's doing, not the gateway's: the reporter's own request bodies, replayed at full size against the vendor's chat and Messages endpoints through the gateway's build and beside it, streamed both ways on the day. Nothing locked the translation's pacing, though, so a future hold would read as the vendor's again. TestChatFromAnthropicStreamsAsItComes paces a Messages upstream at 200ms a word — glm-5.3 at an Anthropic-only provider, and a Claude model where Chat and Anthropic URLs both sit, which the Messages redirect takes there anyway — and wants the client's first chunk before the upstream has written its last word (relative to the upstream, not the wall clock, so a loaded runner's setup time is not held against it), at least two of the five pauses reaching it, and one ask on /v1/messages; holding every event of the chat encoder until finish (as the report read) fails it on both branches with the first chunk at the stream's end. And the keepalive table that walks every protocol's quiet stream (TestQuietStreamsKeptAliveEveryProtocol) had responses, anthropic, chat and anthropic from chat; chat from anthropic, the path a chat client on a Messages-only provider takes, was the row missing, and it keeps alive the same as the rest. go test -race -count=2 on the two tests, go vet and the nogui build pass. (#1310)
- gui: make the Routing context lazy and align wide cards (#1354)
- cli, usage: magpie usage takes --last and --days , a window ending now, with or without --csv and --account (#2075, tinysnake, fcy222fcy)
- agent: a Max output the user set reaches VS Code, Copilot JetBrains, ZCode, WorkBuddy and CodeBuddy as set (#2064)
- gateway: the model's text between a batch's tool calls goes ahead of the batch's first call, so OpenCode Go stops answering "No tool output found" (#2100, eugene1003)
- gui, update: Settings' Restart automatically restarts into a downloaded update once no agent request is in flight (#2096, xiaozhu1337)
- provider: a moved Factory updates to plugin 0.1.26, Gemini's route and a 413 said as Factory's byte limit (#2080, #2082)
- provider: the built-in Factory offers Gemini 4 Argon, as droid 0.241.1 does (#2080)
- gui: an import's results offer Import more… beside Done, so the next batch needs no Done first (#2131, ump45nose)
- gui: Settings › Images says where the Magpie Image tool comes from, turns the Library on, or opens its sheet (#2091, nailuoGG)
- cli: magpie quota refresh [] reads that card again, as its refresh button does (#2104)
- tui: a provider's models view fetches its list again with ctrl+r, and shows what came (#2113, huadi)
- agent: while Claude Desktop is on magpie, Explore inherits Desktop's mythos-magpie model instead of being capped to Opus (#2087, NeuroPaw)
- agent: Claude Desktop on magpie starts from the user's own gateway profile that points at magpie, so their settings carry over (#2035, capiton0)
- gateway: a remote magpie's quota, credit or rate refusal rests only that model, not the whole remote (bigsteven on Discord)
- gateway: a Google account's Code Assist 404 for a model rests that model, not the account (Packing1 on Discord)
- gateway: Claude Code's advisor goes to a relayed Claude upstream under the upstream's own model name, or is left out (JackChou on Discord)
- gateway: a web search's query is masked before it goes to a search API, Kimi Code's or WorkBuddy's search (lc on Discord)
- provider: a Codex token refresh, once sent, is waited for even when its caller gives up
- usage, sessions: Antigravity CLI's (agy's) own calls are counted, read from its transcripts, and a step it sent through magpie once (#2097, parkavenue9639)
- sessions: a Kimi Code session whose state.json names no folder takes it from session_index.jsonl (#2103, infinitr0us)
- gateway, gui: routing history keeps the last two levels of the folder an agent works in, and By session, the stage and the panel name it (#1068, emo172)
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>
更新
- 网关测试：TestClaudeRunEndsAfterStructuredOutput 会等桥接放开它的运行，而不只是等进程结束
- 提供商：已迁移的 WorkBuddy 用户升级到插件 0.1.20，Claude Code 的工具结果可以传到 WorkBuddy AI 的 gemini-3.8-flash（magpie-community/plugins#113）
- 界面：托盘的额度线更高，并以阴影标出与匀速消耗之间的差距（#2079）
- 访问、界面、命令行：网关密钥的限额可以每 N 小时重置一次，包括 5 小时周期（#2102）
- 智能体：Claude Code 的 fable 档位在其他供应商的模型上也不再设置（#2090）
- 提供商：为提供商全部模型设置的窗口不会大于某个模型自身的窗口（#2112）
- 网关：只有第一个分块带标记时，Gemini 的 标签也不会混进回答（#2098）
- 网关：Gemini 4 及以后的模型会按请求的级别收到思考级别（#2083）
- 界面：网关密钥的账号菜单会说明某个提供商的账号都没选中时仍会被使用，以及如何让密钥不用它（#2077，KKKKKKKEM）
- 网关：远程 magpie 上密钥自身的限额仍会让整个远程休息，#2077 的情形已在两个网关之间测试（#2077，KKKKKKKEM）
- 目录：遵循 models.dev 的推理开关（#1294）
- 网关测试：从 Anthropic Messages 翻译成 Chat 的回复按供应商发送的节奏流式输出，现在有测试保证这一点（#1310）
- 界面：路由页的上下文延迟加载，宽卡片对齐（#1354）
- 命令行、用量：magpie usage 支持 --last <时长> 和 --days <天数>，即截至现在的时间窗口，可与 --csv 和 --account 一起使用（#2075，tinysnake、fcy222fcy）
- 智能体：用户设置的最大输出会如实传到 VS Code、Copilot JetBrains、ZCode、WorkBuddy 和 CodeBuddy（#2064）
- 网关：一批工具调用之间的模型文字放到这批的第一个调用之前，OpenCode Go 不再回答「No tool output found」（#2100，eugene1003）
- 界面、更新：设置里的「自动重启」会在没有进行中的智能体请求时重启到已下载的更新（#2096，xiaozhu1337）
- 提供商：已迁移的 Factory 升级到插件 0.1.26，支持 Gemini 的路由，413 会说明是 Factory 的字节上限（#2080、#2082）
- 提供商：内置 Factory 提供 Gemini 4 Argon，与 droid 0.241.1 一致（#2080）
- 界面：导入结果旁边除了「完成」还有「继续导入…」，导入下一批不必先点完成（#2131，ump45nose）
- 界面：设置 › 图像会说明 Magpie Image 工具从哪里来，可以开启资源库，或打开其设置面板（#2091，nailuoGG）
- 命令行：magpie quota refresh <提供商> [<账号>] 重新读取该卡片，与刷新按钮相同（#2104）
- 终端界面：提供商的模型视图按 ctrl+r 重新获取列表，并显示获取到的内容（#2113，huadi）
- 智能体：Claude Desktop 使用 magpie 时，Explore 继承 Desktop 的 mythos-magpie 模型，而不是被限制为 Opus（#2087，NeuroPaw）
- 智能体：Claude Desktop 使用 magpie 时，从用户自己指向 magpie 的网关配置开始，原有设置得以保留（#2035，capiton0）
- 网关：远程 magpie 的额度、余额或速率拒绝只让该模型休息，而不是整个远程（Discord 上的 bigsteven）
- 网关：Google 账号的 Code Assist 对某个模型返回 404 时只让该模型休息，而不是整个账号（Discord 上的 Packing1）
- 网关：Claude Code 的 advisor 以上游自己的模型名发往中转的 Claude 上游，否则不发送（Discord 上的 JackChou）
- 网关：网页搜索的查询在发往搜索 API（Kimi Code 或 WorkBuddy 的搜索）前会做脱敏（Discord 上的 lc）
- 提供商：Codex 令牌刷新一旦发出，即使调用方放弃也会等它完成
- 用量、会话：统计 Antigravity CLI（agy）自身的调用，从其对话记录读取，经过 magpie 的步骤只计一次（#2097，parkavenue9639）
- 会话：state.json 没写文件夹的 Kimi Code 会话，从 session_index.jsonl 取得文件夹（#2103，infinitr0us）
- 网关、界面：路由历史保留智能体工作目录的最后两级，「按会话」、舞台和面板都会显示（#1068，emo172）
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/v0.1.1233