Is Routed Local Routing Suitable for Prototyping?

Recommended by a friend, I tried Routed to see how useful it is. It's a local hybrid router. Simply put, it routes traffic for digital assistant skill files. When a request comes in, it decides whether to hand it to an on-device small model, a local GPU, or a cloud API. My understanding of 'zero token' is that routing decisions try to minimize consuming external large model tokens, making it suitable for pushing down light tasks like repetitive classification and skill selection.

I ran two tasks through it. One was customer feedback classification, requiring sorting "bill unclear," "login slow," and "want PDF export" into billing, performance, and feature categories. The other was reading a React component description and generating a summary for non-technical colleagues.

The former leans on rules, the latter on language understanding. I mainly compared it with WorkBuddy. I've used WorkBuddy for less than a week; its strength lies in intuitive collaboration spaces where multi-Agent task boundaries are laid out. Routed, on the other hand, acts more like a router with a thin interface, relying almost entirely on configuration and logs to understand why it chose a specific path.

The first installation had some hiccups. The README covered installation, but I missed how to configure the local model path. Logs only reported 'model not found.' It took several minutes checking the README to align the paths. But once running, it was indeed lightweight. Official docs claim under 20ms; I didn't stress-test rigorously, but it felt very fast. Classification tasks returned results almost instantly, with logs showing it hit keywords first, then fell back to a lightweight model. The React summary wasn't ideal—it dumped technical jargon directly into the summary, still too harsh for non-tech colleagues. This seems more like insufficient downstream model capability; Routed didn't pass down the constraint of the output audience.

From an engineering perspective, what I like about Routed is transparency. It breaks routing decisions into checkable steps, which helps teams troubleshoot. I previously wrote about multi-Agent collaboration, and my view hasn't changed: the key to collaboration is whether task boundaries can be accepted/verified. Routed exposes these boundaries in light routing but also reveals shortcomings: no team permissions, auditing, version rollback, nor does it turn failure cases into reusable rules. Great for personal prototyping, but I'd question deploying it directly to production for multinational teams.

Currently, I use it for local experiments, cost-sensitive small tasks, and judging whether to go to the cloud. I won't plug it directly into team production workflows. Saving tokens is superficial; whether new hires can understand how requests are routed and attribute errors determines whether to continue investing. If local routers like this improve logging standards, they might become foundational components of agent platforms.

---

📌 This article is compiled from Hacker News. Original text: https://github.com/bshea-1/Routed

Copyright belongs to the original authors. This is a compilation and independent analysis based on public reports.