Initially started as WhatsApp Team Inbox. But, now I am adding telegram bot, emails and MCP support. Eventually, I am going to add like Messenger, Instagram, Line, Viber, Discord, Slack, etc...

So, it will be a bridge between your/your business communication and AI via MCP. If you want to create your own agent you would be able to create as well.

I vibe coded mostly, but once a while I go in and refactor heavy.

It's pretty simple to set up and is designed to "just work", set and forget.

Recently, I've been experimenting with an Elixir Lua runtime (https://deflua.com/) to build custom feeds by parsing HTML, which has worked exceedingly well, so hoping to release that soon.

It's an extremely niche product, but has been growing well organically, which is nice for a side project!

https://news.ycombinator.com/item?id=49686888 https://agentspork.com/

It started with a family member telling me they were frustrated with existing websites. What I noticed is that in many cases they feel cluttered and not smooth to use, almost like a 1-to-1 translation of the underlying database queries, where they have you choose filters and sorting options upfront then browse the paginated results.

I wanted to avoid all of that by creating a different experience; I built an interactive full-screen wall UI with infinite browsing so that most of the screen space is dedicated to the core content. I worked on the name sampling algorithm to try to make it as enjoyable to browse as I could, and the only filter displayed by default is the one to select genders.

More advanced queries are available through the search field which supports natural language queries, running on a local multilingual interpretation engine with a cloud LLM fallback. The database engine is custom built with Rust, compiled to WASM for higher performance when running on CloudFlare Workers. I don’t use an external DB service for names today.

What I’m pretty proud of is to know a lot of technical work is hidden behind that UI, and how it has been shaped by the experience I wanted instead of the other way around.

I am building a new kind of LLM runner which lets you understand the internal reasoning process, and which can be automated to intercept hallucinations and prompt injections before they cause more damage.

(No public code yet, but I do intend to open source it)