I started working on something similar to OpenClaw on Dec. But I'm not the type of guy that is local only. I wanted flexibility. So I built it in a way that lets you decide where you want to run the session. Local or remote. It was an app that included some unnecessary heavy frontend, and I got distracted taking care of things no one cares about.

We had a trip to Japan planned and a week before our departure I had the unhealthy hope that it'll all work before we leave, and I can have my own autonomous company making me money as we travel. That week was a grind and, long story short - it didn't work, and I spent a few days somewhat disappointed, instead of enjoying Tokyo. But Tokyo is Tokyo and two kids don't really give you that feeling to stick around. I let it be.

Then, we fell in love with Japan. Mostly the less known parts. 5 weeks passed quickly and, the flight date was getting closer.

Out of nowhere, I realized that the thing that was the most buggy was the session. It wasn't just one session. I had 3 types of sessions, it all went to diff compute, and I just had a strong intuition that I have to get the session right..

I decided to focus on solid, reliable sessions. I just wanted the thing to work.

This is where Cerver came to be. It's an effort to create a reliable session infrastructure. A session is basically a (1) transcript (2) compute (3 harness and (4) model

So Cerver is an API, that lets you control 1, 2 3 and 4 in a single call. You can swap all. Swap Compute, swap harness. let them consult each other, chain them, make them run in parallel local or remote.

It can be powerful, and also simple to use. Sometimes Claude is just not delivering and I say to Claude - "hi, please consult codex about this." and it really works. They complete each other.

Also supports Gemma, Ollama, GLM.

Would love your feedback. Eyal cerver.ai