Sweetheart is the AI agent in the AppSweet desktop app. This page describes the current alpha desktop app and the upcoming Local runtime, marked below. The last section covers longer-term plans.
What it does in the alpha
- Chat. Ask in plain language and get a streamed reply. You can stop a reply, retry it, and come back to earlier conversations.
- Todo proposals. Attach a todo list and Sweetheart can read it and propose up to 20 new or changed todos: titles, descriptions, due dates and done state. Nothing changes until you apply the changes, and suggested due dates are confirmed separately. You can also reject the proposal or request revisions.
- No Docs access yet. Sweetheart cannot read or write your Docs.
Hosted models
The current public alpha (desktop build 39) supports OpenAI. In Sweetheart's setup, create or unlock your vault, save your OpenAI API key and model ID, and allow the profile for the current organisation in the provider settings.
For Todo proposals, use the exact model ID gpt-4.1-2025-04-14. Other OpenAI model
IDs can be configured for chat, but do not inherit Todo tool support. Availability
also depends on your OpenAI account and the organisation's AI policy.
Anthropic and Google support is implemented in development but is not included in this public desktop download. Check the release notes before choosing a build for those providers.
Provider keys stay on your Mac, in Sweetheart's encrypted vault, and are never sent to your AppSweet server.
Local models
Local models are implemented in development but are not yet included in the public alpha download. The current download uses hosted providers. The models below describe the upcoming Local runtime; check the release notes before choosing a build for local inference.
The upcoming Local runtime can download a model and run it on your Mac. Prompts and replies then never leave the machine. Every local model can make Todo proposals.
| Model | Mac memory |
|---|---|
| Qwen3.5 4B, Gemma 4 E4B | 16 GB |
| Gemma 4 12B, gpt-oss 20B | 32 GB |
| Gemma 4 26B-A4B, gpt-oss 120B | 128 GB |
Downloads range from under 3 GB for the smallest model to about 63 GB for gpt-oss 120B.
Where your prompts go
With a local model, prompts stay on your Mac. With a hosted model, prompts and the todos you attach go from the desktop app directly to the provider you chose.
Before each hosted request, the app asks your AppSweet server whether AI is allowed in the organisation. The server receives the provider, the model name and a fingerprint of the request, not the prompt itself. The server runs no models, so it needs no GPU and has no AI settings in its environment.
What comes next
The items below are the direction of travel, not features of the alpha.
- One request across apps. Sweetheart writing a plan in Docs and filing the todos for it, with every change shown as a card you can open and review.
- More of your workspace. Reading and working with Docs, Goals and the other apps.