The 9Router proxy is a local, OpenAI-compatible gateway that runs on your own machine: your coding tools point at http://localhost:20128/v1, and the router decides which of 40-plus providers actually serves each request. There is no hosted service sitting in the middle, no markup on your usage, and no new accounts to create — it installs into a folder, runs on localhost, and quietly becomes the one address every AI tool on your Mac talks to. After weeks of running my agents through it, it has earned a permanent place in my stack.
📺 Watch: 9Router: New FREE Unlimited AI Coder!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
This article is my deep dive on the proxy layer specifically — the endpoint it exposes, the fallback logic that picks providers, the token compression that saves real money, and the honest caveats. If you want the wider verdict first, my full 9Router review explains the two-router redundancy pattern I run it inside, and my 9Router install guide takes you from an empty folder to a working dashboard.
What a Local AI Proxy Actually Is
A proxy is just one consistent address standing in front of many inconsistent ones. Instead of your coding agent needing to know about Anthropic, OpenAI, Gemini and twenty other providers — each with its own keys, quotas and quirks — it knows about exactly one thing: a URL on your own machine. 9Router listens at that address, receives every request your tools make, and forwards each one to whichever provider it judges best at that moment. Your tool never notices; it simply gets an answer back in the shape it expected.
The word doing the heavy lifting here is local. 9Router installs into a folder and runs on localhost, which means your prompts and your code context travel from your tool to a process on the same machine before heading out to a provider you already trust. There is no third-party relay reading your traffic, no hosted dashboard holding your credentials, and nothing to fall over at 2am that is not already sitting on your desk. When I counted what my own install could see, it was 573-plus models across 40-plus providers — all fronted by one boring local address.
The OpenAI-Compatible Surface: One Endpoint, Nearly Every Tool
Here is the documented shape of the thing. The base endpoint is http://localhost:20128/v1, chat requests go to /v1/chat/completions, and authentication is a Bearer token you generate in the 9Router dashboard. That is the entire integration story, and it is deliberately dull — it copies the shape of OpenAI's own API, the industry's de facto standard.
That compatibility is the strategic part. Nearly every CLI agent, IDE plugin and agent framework shipped in the past two years exposes two settings: a base URL and an API key. Point the base URL at the local endpoint, paste the dashboard token in as the key, and the tool routes through 9Router without knowing anything changed. No SDK swaps, no plugins, no forked configs. My how to use 9Router walkthrough covers the general pattern, and my 9Router with Claude Code guide covers the specific tool I point at it most often.
Want the exact router configs I run behind my content agents — providers, fallback order, token settings, the lot? Join AI Profit Lab and copy my local proxy stack directly, or book a free SEO strategy session and I will map it onto your business live.
The Routing Brain: Smart 3-Tier Fallback
Forwarding requests is easy; deciding where to forward them is where 9Router earns its keep. The documented system is called Smart 3-Tier Fallback, and the order is simple:
- Subscription first. Requests go to the plans you already pay for — the repo lists five subscription providers: Claude Code, Codex, GitHub Copilot, Cursor and Antigravity.
- Cheap second. When subscription quotas run dry, the router drops to low-cost API-key providers — there are 20-plus of those wired in.
- Free third. As a last resort it falls to the three free-tier providers: Kiro, OpenCode Free and Vertex AI.
The switching is automatic. When a quota runs out or a provider starts throwing errors, 9Router moves to the next option without your tool ever seeing a failure — the repo's stated design goal is zero downtime. In practice, an agent that would have stopped dead at a quota wall instead gets slightly cheaper and keeps working, which for long unattended jobs justifies the install on its own.
RTK Token Saver: The Compression Layer That Pays Rent
The second documented feature most people miss is RTK Token Saver. Coding agents are wasteful by nature: every git diff, grep result, ls listing and tree output gets stuffed into the prompt and shipped to the model, and agents re-read that context constantly. RTK compresses exactly those tool outputs before they are sent, and the repo puts the saving at 20 to 40 percent of input tokens per request.
It is enabled by default, and if a specific request needs the raw output you can switch it off per request with the header X-9Router-Token-Saver: off. Twenty to forty percent sounds abstract until you remember agent workloads are dominated by input tokens — I broke down those economics in my Hermes memory budget piece, because context re-reading is where the money actually goes. Compression applied to the noisiest part of every request is a saving that compounds without you ever thinking about it.
The OAuth Trick: Routing Through Access You Already Own
This is the part that made me sit up when I first tested 9Router. The subscription tier does not ask you to create new accounts or buy new API credit — it wires into the OAuth sessions you already have. Signed in to Claude? Codex? GitHub Copilot? Cursor? The router routes requests through that existing, already-paid-for access. You are not buying anything new — you are finally using what you already own.
Most people I speak to pay for two or three coding subscriptions and use a fraction of each; fronting them with one proxy means quota you already bought gets consumed first, before a single penny of fresh API spend. That is the quiet financial argument for the whole tool.
Honest Notes on the 9Router Proxy
A few caveats, because no tool deserves a clean sheet. First, free tiers move — the README carries its own warnings about that, and any given free provider could tighten limits or vanish overnight. Second, the cost figures in the dashboard are estimates for tracking, not bills: the router itself never charges you anything, so treat the numbers as directional rather than accounting-grade. Third, it is MIT-licensed open source, so check the code before trusting it with your workflow — it lives at decolua/9router on GitHub, and I picked apart the repository itself in my 9Router GitHub breakdown. None of these are dealbreakers; they are just the difference between using a tool and understanding it.
Where the Proxy Fits in My Stack
In my own setup, 9Router sits underneath the free-coder tab of Agent OS — the layer where routine drafting and grunt work should cost as close to nothing as possible. The proxy handles the plumbing question of which provider answers; the separate question of which routed brain actually wins real jobs is what Goldie Bench exists to answer, because routing to a cheap model that produces rubbish is a false economy.
The Proxy at a Glance
| Proxy piece | Documented detail |
|---|---|
| Base endpoint | http://localhost:20128/v1 — served locally on your machine |
| Chat route | /v1/chat/completions |
| Authentication | Bearer token generated in the dashboard |
| Fallback order | Subscription, then Cheap, then Free — automatic switching on quota exhaustion or errors |
| Providers | 5 subscription, 3 free-tier, 20-plus API-key providers |
| RTK Token Saver | Compresses tool outputs (git diff, grep, ls, tree and similar); saves 20-40 percent of input tokens; on by default |
| Saver override | X-9Router-Token-Saver: off — disables compression for one request |
| Cost figures | Estimates for tracking only — the router never charges |
| Licence | MIT, open source |
Frequently Asked Questions
What is the 9Router proxy?
The 9Router proxy is a local, OpenAI-compatible gateway that runs on your own machine and sits between your AI tools and 40-plus providers. Your tools send every request to one localhost address, and the router picks the provider — subscriptions first, cheap API keys second, free tiers last — switching automatically when quotas run out.
What is the endpoint?
The documented base endpoint is http://localhost:20128/v1, with chat completions served at /v1/chat/completions. Requests authenticate with a Bearer token you generate in the dashboard.
Does it work with any OpenAI-compatible tool?
That is the design. Any tool that lets you set an OpenAI-style base URL and API key can route through 9Router — which in practice covers almost every modern coding agent, CLI and IDE integration.
What is RTK Token Saver?
RTK Token Saver is 9Router's documented compression layer. It shrinks tool outputs such as git diff, grep, ls and tree before they reach the model, saving 20-40 percent of input tokens per request. It is on by default and can be disabled per request with the header X-9Router-Token-Saver: off.
Is my traffic going through someone else's server?
No. 9Router installs into a folder and runs on localhost, so requests go from your tool to a process on your own machine and then straight to the provider. There is no hosted middleman, and the router itself never charges you.
My Verdict
The 9Router proxy does the least glamorous job in an AI stack — answering one local URL and deciding who picks up — and does it well enough that I stopped thinking about providers altogether. One endpoint, three fallback tiers, default-on compression that trims 20-40 percent off the noisiest tokens, and all of it running on your own machine through access you already pay for. It will not make your output better on its own; it makes everything underneath cheaper and steadier, which is precisely what infrastructure is for.
If you want my full playbook for running agents on routed, nearly-free models — proxy included — join AI Profit Lab and get the complete routing setup I use daily, or book a free SEO strategy session and we will build your version of it together.
