下载 Draggy

系统要求

	最低配置	推荐配置
系统	Windows 10 或 11，64 位	Windows 10 或 11，64 位
处理器	64 位 x86 处理器	64 位 x86 处理器
内存	8 GB	16 GB，或 8 GB 显存
显卡	不需要。模型在处理器上运行。	NVIDIA（CUDA）、AMD（Vulkan 或 ROCm）或其他支持 Vulkan 的显卡
存储空间	5 GB 可用空间	10 GB 可用空间
网络	需要联网，从 Hugging Face 下载模型。	需要联网，从 Hugging Face 下载模型。

运行下载的安装程序。
SmartScreen 会拦下安装程序。请选择更多信息，然后选择仍要运行。
Draggy 会带你完成设置。它检查你的硬件，并推荐一个合适的模型。在模型这一步继续，即可从 Hugging Face 下载该模型。你可以选择更小或更大的模型，也可以改为连接提供方。

解压下载的文件，把 Draggy 移到“应用程序”文件夹。
首次打开会被拦截。打开系统设置 → 隐私与安全性 → 仍要打开，或在终端运行 xattr -dr com.apple.quarantine /Applications/Draggy.app。macOS 版本无法自行更新。
Draggy 会带你完成设置。它检查你的硬件，并推荐一个合适的模型。在模型这一步继续，即可从 Hugging Face 下载该模型。你可以选择更小或更大的模型，也可以改为连接提供方。

	最低配置	推荐配置
系统	64 位 Linux，带 GTK 4.14、libadwaita 1.5 和 WebKitGTK 6.0，例如 Ubuntu 24.04 或 Fedora 40	64 位 Linux，带 GTK 4.14、libadwaita 1.5 和 WebKitGTK 6.0，例如 Ubuntu 24.04 或 Fedora 40
处理器	64 位 x86 处理器	64 位 x86 处理器
内存	8 GB	16 GB，或 8 GB 显存
显卡	不需要。模型在处理器上运行。	NVIDIA（CUDA）、AMD（Vulkan 或 ROCm）或其他支持 Vulkan 的显卡
存储空间	5 GB 可用空间	10 GB 可用空间
网络	需要联网，从 Hugging Face 下载模型。	需要联网，从 Hugging Face 下载模型。

在 Debian 或 Ubuntu 上，从下载目录安装 .deb：
```
sudo apt install ./draggy_3.0.9-1_amd64.deb
```

在 Fedora 上，改为安装 .rpm：

sudo dnf install ./draggy-3.0.9-1.x86_64.rpm

Draggy 会带你完成设置。它检查你的硬件，并推荐一个合适的模型。在模型这一步继续，即可从 Hugging Face 下载该模型。你可以选择更小或更大的模型，也可以改为连接提供方。

Windows 版本将通过 SignPath Foundation 进行签名，该基金会为开源项目免费提供代码签名。

Commands in Code mode can keep running as background terminal tasks while you keep chatting
The statistics tab is rebuilt, with a cost section, budgets and editable prices
Image generation accepts reference images
Provider model lists refresh on their own
AMD graphics cards can use the ROCm build of llama.cpp, switched on under Settings, Models (off by default)
The model list shows each model's format
Embedding models show an icon instead of a second tag
The message queue card is easier to read and use

Interactive question cards let models ask for choices, multiple selections, or free-form text input directly in chat
A question session bar displays questionnaire progress with next, submit, and skip controls
Question history cards allow expanding and reviewing previous answers
Settings under Chat Preferences configure question timeouts, auto-skip behavior, and toggling interactive questions
Image attachments and pasted images in the composer display thumbnail preview cards with one-click removal
Skills can be invoked through slash commands in both Chat and Code modes
Native macOS and Linux builds reach parity with Windows for file browsing, diff reviews, terminal tabs, and canvas tools
Streamlined tool execution by running code directly through project terminal commands

Removed the unreliable explore subagent in favor of direct search and file reading tools
Prevented star notice popups from appearing during video recordings and demos
Updated project dependencies and action workflows

macOS and Linux builds are back in the automated release workflow alongside Windows, producing native packages on each release
An embedded terminal pane opens in the dock with shortcuts Ctrl+Shift+T on Windows or Linux and Cmd+Shift+T on macOS, supporting multiple tabs and custom shells
Code blocks in chat have a button to run commands directly in the terminal
The Diff pane adds History and Branches tabs to inspect commits and switch local branches
Added shortcuts to open the Terminal with Ctrl+Shift+T, open the Plan with Ctrl+Shift+L, and close the active dock pane with Ctrl+\
Chats can be exported to Markdown or standalone HTML from the sidebar menu, including thinking steps, tool activity, diffs, and metrics
Follow-up messages can be queued while a reply generates, with Enter to queue, Ctrl+Enter to send next turn, and Ctrl+Shift+Enter to send immediately; queued messages persist across restarts
When a reply reaches the context limit and the model has room, Draggy automatically expands the window to the next size bucket and continues generating
Interrupted replies can be resumed if the app closes during generation
Error notices specify whether a failure came from the model engine, memory limits, network drops, or rate limits
Work summary blocks show counts of files created, moved, and deleted
Added provider icons for Nebius, Novita, SambaNova, SiliconFlow, and custom endpoints

Fixed an issue where conversations could reset or drop messages during turn transitions
Rewrote the diff parser to handle multi-file diffs, untracked files, and mixed line endings cleanly
Fixed context meter totals on multi-pass agent turns to avoid double-counting earlier passes
Fixed race conditions when stopping a generation turn
Updated bundled Electron, undici, and fast-uri dependencies