About five months ago, I started thinking seriously about how we run agents — and how little control end users actually have over privacy, data, and compute. Running an LLM locally is great for rewriting emails or answering basic questions, but it doesn’t give you real autonomy. What does give you power is the ability to run full-fledged local agents capable of deep research, coding, reasoning, and tool use — all without sending data to external services.
With that vision in mind, I started an open-source project (MIT + Apache-2 licensed) to make building and running agents modular, customizable, and hardware-friendly. Every component of the agent stack can be swapped out:
Memory: plug in your own memory backend
LLM layer: use local models or cloud APIs interchangeably
Execution style: choose between ReAct, CoT, or your own custom executor
The goal is to enable real research workflows while maintaining strong performance and developer-friendly usability.
The project doesn’t stop at building agents, either — we want to ensure you can run them efficiently on any accelerated hardware, from GPUs to edge devices.
Would love feedback from the HN community!