Your AI agents don't all need to run on the same machine
AI agents don't all have to be crammed onto one machine. NAISYS centrally manages hosts, model accounts, and tasks, with scheduling and monitoring in the browser. Open source under MIT.
Your AI agents don't all need to run on the same machine.
Dan Kornas released an open-source project called NAISYS. It is a self-hosted system that lets teams unify compute from multiple machines, model accounts, and tasks into a shared organization used by both people and AI agents.
The architecture is straightforward: each host connects to a central hub, and agents, permissions, and running tasks are managed through a browser UI.
Key capabilities:
- Assign tasks by host: agents can run across the entire cluster, or be pinned to machines with the right GPU, operating system, or desktop session
- Real-time browser monitoring: consoles, run logs, screenshots, and host status all in one view
- Independent budget per agent: set rolling spending caps and automatically pause when the limit is reached
- Hosted and local models: connect OpenAI, Anthropic, and Google APIs, and run local LLMs too
- Complete agent toolset: shell, browser, desktop operations, and inter-agent email
Open source, MIT license. The code is on GitHub: https://github.com/swax/NAISYS
What makes this approach interesting is that it acknowledges a reality: individuals and teams often have more than one machine—laptops, GPU boxes, idle servers—each with some compute, but it is hard to put that compute to use when running AI agents. NAISYS treats this as a resource-scheduling problem.
People in the comments are already asking key questions: When two agents share one model account, how should rate limits be split? What if a task needs to access files on another machine? How do you keep agents from stepping on each other's work?
These questions do not have answers yet, but they point to where this kind of system is genuinely difficult: pooling compute may only be the beginning; file synchronization, permission isolation, and concurrency coordination are the real headaches, and the more machines there are, the more obvious it gets.
For teams with a few machines that want to use agents for development, it is worth a try.
发布时间: 2026-10-27 15:03