-
Notifications
You must be signed in to change notification settings - Fork 0
Settings
Nine panels, in the order they appear.
Light or dark theme, and three text sizes for the chat. The theme is applied everywhere, including the voice orb, which is dark on light and white on dark.
The model that answers you, and every model you have installed. Each is shown with its size and how it will fit: the panel estimates how much lands in VRAM and how much spills to the processor, which is the number that predicts whether a model will feel usable.
You can search Ollama's catalogue from here and pull something new, or delete a model you have finished with. Embedding models are listed but cannot be chosen, because they cannot generate a reply at all — Draggy asks Ollama what each model can do rather than guessing from its name. Every other model is selectable, however small.
Indexing your own documents. See Document Library.
Web search mode — auto, on or off — and the search provider: Automatic, Brave Search, DuckDuckGo, Startpage, Brave Search API or SearXNG, plus the URL and API key fields the last two need.
Code execution lives here too, off by default. See Tools and Capabilities.
The ad blocker is deliberately not here. Its switch is in the toolbar of the browser window itself, which is where you find out you want it — see Browser and Ad Blocking.
The MCP server catalogue: thirty-four servers that lend the assistant tools it does not have on its own, none of them duplicating a provider Draggy already ships. Everything is off until you turn it on, because turning one on runs somebody else's program on your machine. See Extensions.
The talk model and the embedding model are picked here — see Choosing a Model and Document Library. The talk voice lives here too: system or natural, which voice, and speed.
Custom instructions, added to every system prompt. Short and concrete works far better than long and aspirational, particularly on a small model: "answer in British English" and "do not use bullet points unless I ask" are the kind of thing that survives.
Thinking mode is here as well — Fast, Balanced or Deep, and on the toolbar of the chat itself if you would rather change it mid-conversation.
Balanced and Deep ask the model to reason before it answers: on a model with
native reasoning this maps to its own setting, and on one without it asks for a
<think> block, which Draggy hides from the reply and shows as a collapsible
section above it. Deep is genuinely slower and only pays off on problems that
deserve it.
Fast turns reasoning off rather than asking for less of it. A model that supports the switch is told not to reason; one that does not is asked, in the system prompt, to answer immediately. Saying nothing was not enough — a reasoning model with no instruction either way reasons anyway, which is how a one-word reply used to cost seven seconds of thinking first.
Twelve interface languages: English, French, Spanish, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean and Arabic.
This changes Draggy, not the model. Whether the model answers you in Portuguese depends on the model. It also sets the language the speech recogniser expects in voice mode, which does measurably affect transcription quality, so it is worth setting correctly even if you read English fine.
How much you have stored — chats, messages, attachments and their size — plus buttons to clear conversations and to open the log file when something needs diagnosing.
Chats live in SQLite in the Draggy data folder. Deleting them is immediate and there is no copy anywhere else.
The current version, the Automatic updates switch, and the update status with manual check, download and install buttons. See Installation for what automatic actually does.