Meet Vi
A personal assistant powered by AI.
If you opened this link through an email, then that email was actually send by my assistant, Vi. This article describes its on-going development.
When I played Horizon Forbidden West, I liked the idea of having a super intelligent entity with capabilities of controlling massive aspects (of the game's reality). Many examples exist of such intelligent systems (Gravis in Iron Man, Baymax in Big Hero 6, etc.), but what was interesting Gaia was the fact it the ability to learn and gain new abilities (although the player had to collect things for that to happen). After the recent developments in LLMs and LLM agents, I started invisioning how a system like this would be implemented.
Claude Code (as if you didn't hear enough about it). I have been using it ever since it was released to the public in May 2025. I can confidently say I have an idea of its inner works. And with the release of channels, the implementation was made easier (still possible without, but more tedious). So I decided to use Claude Code as the harness for this system, which I dupped Vi, create a messaging channel, and make a repository that Vi can edit and added an automation system by which edits to the repository triggers deployment (whether its made by Vi or me from another computer).
The system currently works great so far. I added access to a google account so that Vi can manage tasks, calenders, and emails. Then I had Vi create a bunch of services to track tasks, diet, gym, weight, and projects, and it was able to make those changes concurrently with me working on Cron feature to replace the cron tools provided by Claude Code (they can only last 7 days; I made them indefinite).
Security and Privacy
One of the biggest concerns regarding this implementation is the fact I'm giving Vi credentials of many services that could be a problem if a malicious hacker gained access to, which is why Vi is configured to, in addition to access keys, only respond to registered devices. Vi continuously scrutinize its structure to check for any vulnerabilities, but that's only for the programatic access. For prompt injection related vulnerabilities, in addition to the LLM's own resilence against PI, most of the keys are protected such that Vi can use them but never be able to read them. So in a hypothetical scenario where a hacker can interact with Vi and try to manipulate it, the damage is inheritly limited. Of course, it will never be zero, but I actively try to reach there.
As for privacy, my own privacy I don't care much, but the obvious concern here is having an LLM read emails, mainly if that LLM is provided by an online service. I agree with this concern, which is why my LLM backend is a local one (yes, claude code supports changing backends, but it still expects the Messages API). Main helper here is LiteLLM, which can morph protocols (OpenAI, Anthropic) to others.
Still, I wouldn't consider this a finished work. There are so many things I want to add, and the main concern with any feature is always security; otherwise, you would see Vi reading my SMS messages (spoiler).