I’ve made Ur-Agent-Team open-source today: https://github.com/uncle-tyson/Ur-Agent-Team
It’s a local-based Agent Hub that coordinates AI agents operating on a single machine.
The core idea isn’t tied to any one vendor or integrated development environment. Any connected agent can act as the “brain” of the system, assigning tasks to other agents, receiving their responses, and continuing the workflow, with humans still in control of the process.
Codex, Claude Code, and Antigravity are the first workflows I’ve been testing. But they’re just examples, not the ultimate goal of this project. The ultimate goal is to create a hub that’s easy for various tools to integrate with. This way, other local agents, CLI tools, IDE interfaces, and MCP-compatible tools can all work together within the same coordination framework.
What it focuses on:
- Local-first orchestration, rather than using a hosted coordinator
- Any-agent-as-brain delegation
- MCP server and CLI entry points
- Session-aware dispatch, ensuring that tasks are assigned to the appropriate conversation.
- Reply-return: This allows the orchestrator to verify what response was actually given by another agent.
- A live Agent Hub panel that provides insight into handoffs and replies
– A durable daemon for handling long-running tasks, ensuring that these tasks continue to run even after the editor is restarted. - Explicit safeguards to prevent silently hijacking the wrong session
Why I built it:
In practice, real AI coding workflows already involve multiple agents working together. One tool might be best for handling command-line tasks, another for working within an IDE, and yet another for tasks like code review, research, or running long-running background processes. I wanted a simple, local system that allowed these various agents to cooperate with each other, without forcing everything to be handled through a single tool.
This is still alpha software, and its development was driven by real-world workflow challenges. The current adapters represent the tested solutions, and more hosts will require adapter support. But the core functionality of the software is now useful enough to be shared with others.
If you’re experimenting with local multi-agent coding workflows, I’d love to hear your feedback, any issues you encounter, or any PRs you might have.