Building an AI Agent Workstation in Nairobi: Integrating Open-Weight Models for Agentic Workflows

This article details the creation of Atlarix, a lightweight 400MB AI agent workstation designed to operate alongside IDEs like VS Code and IntelliJ. Developed solo in Nairobi by Amariah Kamau under NorahLabs, the system prioritizes open-weight models such as DeepSeek, Qwen, Kimi, and MiniMax for agentic tasks while maintaining BYOK compatibility for other tools.

What I Built: Atlarix Workstation

Atlarix is a compact AI agent workstation engineered to coexist with integrated development environments (IDEs) rather than replace them. Its 400MB footprint ensures minimal system resource consumption while focusing on open-weight models like DeepSeek, Qwen, Kimi, and MiniMax. These models are positioned as core components of the system, with BYOK (Bring Your Own Kit) functionality retained for non-specified workflows.

The workstation was developed entirely in Nairobi, Kenya, under the NorahLabs initiative. This localized approach emphasizes accessibility and practical deployment for real-world agentic applications. Unlike traditional setups that bundle models as afterthoughts, Atlarix integrates these open-weight solutions as primary tools for agentic reasoning and task execution.

The Problem: Gaps in Existing Solutions

The primary challenge addressed by Atlarix was the inefficiency of existing frameworks for running open-weight models in agentic contexts. Many solutions either replaced IDEs entirely or treated open-weight models as supplementary features rather than central components. This led to fragmented workflows and limited adaptability for developers working on multi-step agentic tasks.

Additionally, the lack of a unified platform for models like DeepSeek, Qwen, Kimi, and MiniMax created barriers to seamless integration. Atlarix resolves this by providing a native harness that unifies these models under a single interface, enabling developers to leverage their strengths without switching between disparate tools.