back
56 comments
We need LLM query routing at the OS level like Mobile data. I know it will sound crazy but hear me out. I think about this AI inference as infrastructure. I do not want to pay for it on every app I use it on. I do not think "I have to pay the mobile data of youtube, and the mobile data of whatsapp etc.". I pay Mobile data infrastructure and let my device route it appropiately. In fact, if we ever go the local llm route, you could have LLM capabilities without having access to the internet (or local LAN), and your OS/computer is the only one capable of doing that routing for you.
It doesn't sound crazy at all, this seems almost obvious. The OS should provide a chat completions server and the user should be able to select the underlying LLM's server. This should be just like selecting a default search engine or browser.

Hopefully the EU forces US tech giants to do this. God knows Apple and Google won't do this on their own. They gotta get that sweet default provider revenue.

Honestly I don't get the point but if you want to explore that, both on desktop, mobile or headless server Linux allows you to try it.

You can run ollama with whatever you want on a Debian in literally minutes. You can even do that within a virtual machine using e.g. QEMU, so that you can do all the tests you need risk free.

Again I don't understand what that would enable that can't be done today but it's perfectly fine, you can try today anyway, no need to ask permission to anyone.

Maybe systemd could do that.
I mean, the reason mobile data is part of the OS is because the antenna is hardware that must be shared across processes. Chat completions is just a network call like anything else—it’s already available to every app; they don’t need to pay separately (they can use the same account), they just pass their API key over the network to the completions server. What am I missing?
It's funny how much that first paragraph is Claude's voice. I don't know how it got trained so hard to use, "the shape of" for everything.
Loads of ed sheeran in the training data?
Thank you! I thought it was just me.
Do you want the honest answer?
Interesting. Read through the comments and the cache issues and context loss issues were already mentioned and they are all good valid points. I'll go from the positive side.

I think the part that resonates most with me is the deterministic rule based routing. It can even double as a policy that you can audit. Configuring which calls would go external and internal.....this could work well for companies who want to enforce certain privacy policies on what data goes to external frontier models vs internal local models. I think the privacy angle is stronger here over the cost angle.

I believe a config-led setup is the way over a model trained on routing. Opus (I'm most familiar with) and I'm sure other models already have the subagent spawning skills down-pat. I basically do a similar routing by config setup but using Claude with some rules and a framework.

So if any routing is needed it might be at the point you've identified....locally on device config-led. Neat. If I get the chance I will check it out and give it a whirl.

I'm not sure I understand what this is trying to solve?

If a prompt I give routes to one model, and then another prompt to another model, how does one tie the context together such that the next model knows what's going on?

Otherwise this would only be useful for one-off prompts as far as I can tell.

And if it did keep a context to be passed around, it would always land hot (not in the cache).

Every turn of a conversation with an LLM is getting the whole conversation. Caching complicates the picture, but not by a huge amount. That's why a short question at the end of a long conversation chews tokens faster than it would in a fresh session.

So, a conversation that's ongoing with one model then switching to another would presumably send the whole conversation and the new question. Which defeats the purpose of splitting traffic...so, you're not wrong to question how this actually improves things for anything other than short sessions, which you could choose your own model for if it's a small problem.

Here's a use case: You want to extend the GPT 5.5 quota in you Codex subscription by routing some % of requests to DeepSeek V4 Pro. A router needs to figure out which requests to route where, for the appropriate difficulty level.

Another use case: You have two models on your local device. One is large and fairly powerful but low, the other is smaller, faster and good at tool calls and chat, but not great for writing and reviewing code. If you route between them per request, you can get a better developer experience with preserved performance.

The linked repo aims to help you achieve these things, as do I with the role-model router and protocol that I linked in another comment.

LLMs have no state. There is nothing remembered, nothing new learned. It's the same input , the same output always (unless seeding is randomized). So during a chat it won't matter if every chat turn a different provider is used.
I'm not sure if output of easy commands like "summarize this" are added back to the context? I always assumed they are in a separate UI layer?
Love to see local/cloud routing explicitly supported.

I'm building another router for routing between local and remote models, ShowHN coming up later today. Here's a sneak preview of the github: https://github.com/try-works/role-model

Posted my ShowHN: https://news.ycombinator.com/item?id=48706181
There are so many proxies like this now but I can tell you from first hand experience this is not going to work. You cannot just route away from a situation at such a high level especially when we are talking about models that are quite different in behaviour, with different context windows and tuned to different tool uses. The harness is doing all kind of funky things to compensate for issues (like tool call truncation) that a proxy that routes dynamically like this will work against the very same strategies that make the harness work.

Interesting concept, work in theory, but I cannot see this being part of larger system.

This is not choosing between different models, really. You should check the (interesting, yet sadly very slop-padded) readme. It’s about trying to make a binary decision: Is this a hard or easy question, and about making that decision extremely fast. They suggest putting another router that chooses the model behind it. I’m not sure how well it would work, but the idea is interesting and different than other routers.
we could use some composability.

today any kind of routing requires implementing an http proxy to put in the middle

ideally harnesses would support a routing plugin which receives the new whole context and returns just where to send it, and the harness does that. no http proxy. obviously some complications if you want to route from codex to anthropic or openrouter.

but we need to decouple the context building and routing decision from the actual http requests sending, we need to be able to insert "context/routing plugins" in the chain

Mine does this. Only I don’t use the whole context with the router because that’s wildly resource intensive and slow.

Then, a bit like open router, it does a classifier job with a fast model to choose which one should process the turn.

In my case I usually don’t do local vs remote… although it can. Now I use it for thinking vs no-think against my preferred local model, which is a huge time saver even with the added classification step.

I'm still waiting for an isolated protocol so we don't have to run the hanress directly on any of the code base's infrastructure. Something as simple as piping everything into and out of an ssh shell would be better than anything I've tested so far.
We are developing many applications in my company, some of them safety critical. A natural routing way could happen for certain phases of development, and interfaces via git. One agent works on branch a and is responsible for brainstorm planning specs, and the other is responsible code and tests. The first agent creates tickets for the second one and the second one consumes these. This works with today’s standard harness.
Slight tangent, but “Wayfinder sits behind whatever OpenAI-compatible client you already use” reminds me that descriptions of where proxies sit in the information flow always seem so arbitrary to me:

  - “after the client”
  - “reverse proxy” (in front  of servers)
  - “proxy” (in front of client)
I always have to look this up, surely there must be a standardized way to describe this?
"after the client" and "in front of client" can mean the same thing depending on your viewpoint.
Some kind of routing prompts to different models does make sense. But the usecase of saving money on simple prompt.. I think that has only a slight benefit. Fix my typo doesn't use many tokens anyhow.. also model switching still requires carrying over context so it does have some overhead right.
There's a hidden tax with routing this way, the original model loses context of what was done and either performs a regression or hallucinates.

I think this sort of behaviour started happening more frequently as agentic/ai programming became more often.

Back in the days (lol, reads like a long time ago but that's probably a few months?), you would not say "edit this typo", you would just open the file and not be lazy, and the harness would detect a user change and ground itself.

I feel like now, when I edit outside the AI flow, it goes and introduces a regression or gets lots thinking it didn't do that and something must have gone wrong.

I do this manually with a desktop app called BoltAI that lets you continue the whole conversation at your LLM of choice.
I sure hope the community is working on these APIs now for Linux, with pressure to come on M$ and Apple.
It'd be nice to just have a command prefix e.g.

/local fix my typo

That’s what I did with Pi, super simple :)
Has anyone tried the others listed? Any feedback?
can you send to multiple LLMs to compare responses? From that create a heuristic of which LLM gets what.
This is the way!
I like to think so!
axa
wfdwcZz