What if you could point an AI at your inbox and have it pull out every transaction, bill, and payment notification - without sending a single email to the cloud?
Dwata is a new open-source tool that does exactly that. It connects to Gmail or any IMAP email account, downloads messages to a local SQLite database, and then runs AI models on your machine to identify and extract financial data. The entire pipeline stays on-device.
The workflow has a clever template system. You show the AI a sample financial email - say, a Stripe payment receipt or an electricity bill - and it generates a reusable extraction template. From there, it automatically finds and processes all similar emails in your inbox using that template. Think of it as teaching the AI to recognize a pattern once, then letting it repeat the work across thousands of messages.
Dwata supports multiple AI backends. For fully local operation, it works with Ollama running models like Ministral 3B (tested on a Mac Mini M4 with 16GB RAM). For faster results with a cloud tradeoff, it also supports OpenAI's GPT-4o Nano and Google's Gemini 2.5 Flash Preview. You can switch between models in the settings.
The project is built with Rust on the backend and SolidJS for the browser-based interface, which is a solid technical foundation. That said, the developer is upfront that this is early-stage software. The extraction logic still has gaps, and it won't catch every financial email reliably yet.
The privacy angle is the real draw. Financial emails contain account numbers, payment amounts, and vendor relationships - exactly the kind of data you don't want processed on someone else's servers. Running everything through a local LLM (a language model hosted on your own hardware) keeps that data under your control.