-
Notifications
You must be signed in to change notification settings - Fork 0
Installation
Ollama. Draggy does not ship a model runtime; it drives Ollama over localhost. If Draggy cannot find it on launch it will offer to install it for you, or you can get it yourself from ollama.com.
A graphics card, ideally. Draggy runs on anything, but the experience is decided almost entirely by how much VRAM you have. Below four gigabytes it picks smaller models and leans on the processor, which works but is slow enough that you notice every reply. Apple Silicon does well because the memory is shared with the GPU.
Disk space. Around 10 GB for a typical model, plus a few hundred megabytes for the speech models if you use voice mode. Draggy checks free space before it downloads anything and refuses rather than filling your drive.
Draggy sizes the model to the machine it finds, so nothing here is a hard gate except the disk check. These are the two sets worth knowing.
| System | Windows 10 64-bit (1809 or newer), macOS 11, or 64-bit Linux with glibc 2.28 or newer |
| Processor | Intel Core i5-8250U, AMD Ryzen 3 3200U, Apple M1, or anything newer |
| Memory | 8 GB |
| Graphics | Integrated is fine: Intel UHD 620, AMD Radeon Vega 8. No dedicated card needed |
| Disk | 10 GB free, checked before the first launch |
With no usable graphics memory Draggy picks a model of around 1B parameters and runs it on the processor. A few words a second, and answers plainly less capable than a larger model gives. Every model Draggy picks calls tools, reads images and reasons, but a small one does all three less well, so web search and the browser go wrong more often at this end.
| System | Windows 11, macOS 14, or a current Linux |
| Processor | Intel Core i5-12400, AMD Ryzen 5 5600, or Apple M2 |
| Memory | 16 GB |
| Graphics | 8 GB of VRAM: GeForce RTX 4060, RTX 3070, Radeon RX 7600, Intel Arc A750. Apple M2 with 16 GB of unified memory does the same job |
| Disk | 20 GB free, on an SSD |
Eight gigabytes of VRAM is where Draggy picks Qwen 3.5 9B, the smallest model that holds a conversation, calls tools without needing much patience and answers questions about your own documents well. Apple Silicon does better than its number suggests, because the memory is shared rather than split, and because Draggy hands it MLX builds. See Choosing a Model.
Voice mode downloads a few hundred megabytes of speech models on first use, and the neural voice sounds smoothest with a card that is not already busy generating. The document library adds an embedding model, from 50 MB at the bottom of the ladder to 4.7 GB at the top.
A card with 16 GB, such as a GeForce RTX 4080, an RTX 4060 Ti 16 GB or a Radeon RX 7800 XT, runs a 12B model with voice and the library alongside it. An Apple M3 Pro with 36 GB reaches further, to a 35B mixture-of-experts model. See Choosing a Model for the whole ladder, which goes up to 122B at 96 GB.
Pick your system on draggy.org, which links to the right file for you, or take it from the Releases page yourself.
-
Windows:
draggy_setup.exe. It asks where to install and does not require administrator rights. -
Linux: the AppImage runs as-is once you mark it executable, or install
the
.debwithsudo dpkg -i. -
macOS:
Draggy-<version>-arm64.dmgfor Apple Silicon, or thex64one for an Intel Mac. It is not signed, so see below.
It is not. An unsigned app that was downloaded carries a quarantine flag, and Gatekeeper refuses to open it rather than explaining why. Clear the flag once, after dragging Draggy to Applications:
xattr -cr /Applications/Draggy.appThe same lack of signing means a macOS install cannot update itself, so new versions have to be downloaded from the Releases page by hand. Both go away with an Apple Developer certificate, which has not been bought.
Windows will show a blue SmartScreen dialog the first time you run the installer, saying the publisher is unknown. To get past it: click More info, then Run anyway.
This is worth understanding rather than just clicking through, because the same dialog appears in front of genuinely malicious software.
SmartScreen is not saying it found anything wrong with Draggy. It is saying the installer is not signed with a code signing certificate, so Windows has no publisher identity to attach to it and no history of other people running it safely. Certificates that satisfy it cost a few hundred a year for the kind that clears the warning immediately, and Draggy does not have one yet.
If you would rather verify the download than take that on trust, every release
publishes a latest.yml next to the installer holding its SHA-512, base64
encoded. Compute the same thing in PowerShell and compare the two strings:
[Convert]::ToBase64String([Security.Cryptography.SHA512]::Create().ComputeHash([IO.File]::ReadAllBytes("draggy_setup.exe")))
Or simply build it yourself from source, which is the strongest answer to "how do I know what is in this" and is documented in Building from Source.
The Linux AppImage and .deb are unsigned too, but Linux does not put a dialog
in front of them.
Draggy is not code signed, and that is the whole reason Windows and some antivirus products treat it with suspicion. There is nothing an application can ship that makes a scanner skip it. Anything that did would be a technique for hiding malware, and scanners are built to ignore it.
What actually removes the warning is a code signing certificate, which is a few hundred pounds a year and has not been bought. Until then:
- SmartScreen shows the dialog above on first run. Reputation builds with downloads over time, so it fades on its own.
- Norton, McAfee and similar may quarantine an unsigned installer on sight, or upload it for analysis. That is them doing their job with a file they have never seen before.
- If one flags Draggy, report it as a false positive to that vendor. Every major one has a form, and it is the only thing that clears it for everybody rather than just for you.
- Verify the download yourself with the SHA-512 above rather than taking it on trust, or build from source.
The first launch does more work than the rest. Draggy checks that Ollama is running, looks at your hardware, and if you have no usable model installed it picks one sized to your graphics card and downloads it. That download is the long part: a few gigabytes, once.
The splash screen tells you which stage it is on. If it stalls, see Troubleshooting.
After that, launches are quick. Nothing is downloaded again unless you ask for a different model or turn on voice mode for the first time.
Everything lives in Electron's user data folder for the app:
- Windows:
%APPDATA%\Draggy - macOS:
~/Library/Application Support/Draggy - Linux:
~/.config/Draggy
Inside it are draggy.db (conversations, settings and extension credentials),
library.db (the document index), created_files/, code_runs/,
model-cache/ and mcp-servers/, which is where an extension you switch on is
installed. Models themselves belong to Ollama and live wherever Ollama keeps
them, which is why removing Draggy does not cost you the download again.
Deleting that folder resets Draggy completely without touching your models. Privacy and Data covers what each file holds, how to back it up, and how to uninstall properly.
Draggy updates itself. With Automatic updates on, which it is by default, it looks for a new version twenty seconds after launch and every six hours after that, and downloads what it finds in the background.
When one is ready, the next time you open Draggy a dialog offers Install now or Maybe later. Installing restarts the app and is silent: the installer does not ask where to put it or whether to continue, because you already answered that when you first installed. Choosing later leaves it for the next launch, and it installs on quit in any case.
Turn it off in Settings → Updates and nothing happens unless you press Check
now. The same panel has Download and Restart and install buttons when
an update is waiting.
Updates only work in an installed build. Running from source, the panel will say updates are disabled, which is correct.
Most releases are plain, like 1.2.5. Some carry a word on the end, like
1.2.6-fix. Those are ordinary releases and reach you the same way: the app
does not treat them as a separate channel, and the updater is happy to install
one.
The only rule worth knowing is that a suffix sorts below the number it is
attached to. 1.2.6-fix is newer than 1.2.5, but 1.2.5-fix would be older
than 1.2.5 and would never be offered to anyone already on it.