9Router: Wire Every AI Coding Tool to 40+ Providers, with Automatic Fallback So You Never Hit a Quota Wall

You've subscribed to Claude Pro. Then you bought Cursor. And then your company handed you GitHub Copilot on top of that. So now you're paying for three different pools of AI capacity — and every single one of them manages its quota completely in isolation. When one runs dry, you switch to another by hand. The OAuth tokens expire after a few hours and kick you back to a login screen. And each tool ships its API in a slightly different format. You've paid for plenty of capacity, and yet somehow you keep slamming into rate limits at exactly the moment you need them most.

There's a tool called 9Router that takes a really clever, dead-simple angle on this. It runs an OpenAI-compatible proxy right on your machine. Every AI coding tool you use points its requests at that one proxy, and 9Router automatically routes each request to whichever provider is, at that moment, the cheapest one that still has quota left. It's picked up around eighteen thousand stars on GitHub, it's actively maintained, and it ships as both an npm package and a Docker image.

Let's talk about the core idea, because it's elegant. The traditional setup has every tool configure its own provider and hold its own API key. The moment a provider changes — new key, new endpoint, whatever — you have to go reconfigure every tool, one by one. 9Router pulls that whole layer out of your tools. As long as a tool supports a "custom OpenAI endpoint" — and almost all of them do — you point every one of them at localhost, port twenty-one-twenty-eight. From then on, choosing the provider, switching providers, handling authentication — that's all 9Router's job. Swapping providers becomes 9Router's problem. Your tools get configured exactly once, and you never touch them again.

Installing it is about as simple as it gets: one npm install, run the command, and a dashboard pops open in your browser. Then for each tool — Claude Code, Codex, Cursor, Cline, whatever — you just set the endpoint to that local address, paste in the API key the dashboard gives you, and pick a model. And here's an important detail: your API keys and OAuth tokens are stored only on your own machine. Nothing gets shipped off to some third party.

Now, the heart of the whole system is what they call three-tier fallback. Picture three tiers stacked from cheapest to most expensive. Tier one is your subscription services — Claude Code, Codex, Copilot. The stuff you're already paying a flat monthly fee for, so you want to burn that first. When tier one runs out of quota, 9Router automatically drops down to tier two: the cheap pay-as-you-go APIs, things like GLM at around sixty cents per million tokens, or MiniMax at twenty cents. And if you hit a budget cap there, it falls again to tier three: the genuinely free services. The key word in all of this is automatic. A request comes in, it tries tier one, succeeds, you get your response. Tier one's tapped out? It quietly slides to tier two, and you never notice. If a provider has multiple accounts attached, it'll even round-robin across them. And the dashboard shows you live quota tracking with reset countdowns, so you can actually see where you stand.

On that free tier, three options are holding up well right now. There's Kiro AI, which gives you Claude 4.5 plus GLM-5 plus MiniMax, billed as unlimited free. There's OpenCode Free, which needs no authentication at all and auto-discovers whatever models are available. And there's Vertex AI, which hands you three hundred dollars of free credit to spend on Gemini 3 Pro, GLM-5, and DeepSeek. Worth noting — the README mentions that the free plans for iFlow, Qwen, and the Gemini CLI all got shut down one after another over the course of 2026. So this free landscape shifts. Those three are just the current mainstays.

A couple of the smaller features really tie this together. First, OAuth auto-refresh. For the subscription services that authenticate through OAuth — Claude Code, Codex, GitHub, Cursor, Antigravity — those tokens usually only last a few hours. 9Router refreshes them automatically, before they expire, so a long working session never gets cut off mid-stream and dumps you back at a login prompt.

Second, cross-format translation. Different providers speak different native API dialects — OpenAI's message structure and Claude's message structure, for instance, are just not the same shape. 9Router translates between them on the fly as it routes. So your tool only ever has to speak OpenAI format. You never write a separate adapter for each provider; the proxy handles the conversion.

And third, there's a built-in token compression layer they call RTK, the Result Token Kit. This one targets the contents of tool results — think shell output, the stuff that comes back from a git diff, a grep, an ls, a log dump. It detects what type of output it is, filters out the redundant noise, and only then feeds the trimmed version to the language model. The project claims that saves somewhere between twenty and forty percent of your input tokens per request. Quick heads-up on the name, though — this RTK is a middleware baked into 9Router, and it just happens to share an abbreviation with a totally separate standalone Rust tool also called RTK. Different things.

In terms of reach, it supports basically the whole zoo of CLI tools — Claude Code, Codex, OpenCode, Cursor, Cline, Continue, Copilot, and more. On the provider side, the OAuth ones plus something like forty-odd API-key-based providers. It runs on localhost by default, but you can also launch from source or via Docker — and some folks in the community have even demonstrated deploying it to Hugging Face Spaces as a free, always-on alternative to renting a VPS.

But let's be honest about the catch, because there is one. Using an unofficial proxy to hook into your subscription services is, fundamentally, a way of working around how each vendor intends their product to be used. Individual providers carry a real risk of bans or terms-of-service violations, and that's a tradeoff only you can weigh. And one more honest note: if your actual pain point isn't quota juggling but rather "shell command output is blowing up my context window" — that's a different problem on a different axis, and you might be better served pairing this with a dedicated output-compression tool.

So, three things to take with you. One: 9Router's whole philosophy is to collapse provider management into a single local endpoint — configure your tools once, and let the proxy handle every switch behind the scenes forever. Two: the magic is that three-tier, cheapest-first fallback, plus the quiet machinery around it — auto-refreshing tokens and translating API formats so nothing breaks when it switches. And three: it's genuinely powerful for anyone juggling multiple AI subscriptions, but go in clear-eyed about the terms-of-service gray zone, because that risk is yours to carry.