The vision I had for Safata:
A local agent that runs 100% on your computer.
You can choose the AI provider and model based on your needs - frontier, open source (Chinese, French or American) models
A truly capable agent that gets work done and writes things down.
Humans remain at the center of decision-making, collaborating with the agent to achieve their goals.
You can connect your data (Dropbox, Box, Google Drive, HubSpot, etc.) using API keys.
Everything stays on your computer, with no cloud usage except for model calls. It's also really fast when working with files.
The more you use it, the better it gets, thanks to accumulated context.
So far, for my personal use cases:
It's a complementary agent to Claude Code, allowing me to use OpenAI's GPT models to generate images and ElevenLabs to create music for my apps.
It's great for testing new models with real-world use cases.
I also use it for part of my business workflow, monitoring and fetching relevant LinkedIn and Reddit conversations to gather insights.
Privacy is an attractive point for early adopters. For example, you can also use Mistral's models (sovereign AI use case).
A little anecdote: Safata means "tray" in Catalan, a place where you can put anything you want.
I'm happy to share what I've learned and exchange notes with others working on similar ideas.
victorzhangeu•1h ago
The vision I had for Safata:
A local agent that runs 100% on your computer.
You can choose the AI provider and model based on your needs - frontier, open source (Chinese, French or American) models
A truly capable agent that gets work done and writes things down.
Humans remain at the center of decision-making, collaborating with the agent to achieve their goals.
You can connect your data (Dropbox, Box, Google Drive, HubSpot, etc.) using API keys.
Everything stays on your computer, with no cloud usage except for model calls. It's also really fast when working with files.
The more you use it, the better it gets, thanks to accumulated context.
So far, for my personal use cases:
It's a complementary agent to Claude Code, allowing me to use OpenAI's GPT models to generate images and ElevenLabs to create music for my apps.
It's great for testing new models with real-world use cases.
I also use it for part of my business workflow, monitoring and fetching relevant LinkedIn and Reddit conversations to gather insights.
Privacy is an attractive point for early adopters. For example, you can also use Mistral's models (sovereign AI use case).
A little anecdote: Safata means "tray" in Catalan, a place where you can put anything you want.
I'm happy to share what I've learned and exchange notes with others working on similar ideas.