惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

The GitHub Blog
The GitHub Blog
Martin Fowler
Martin Fowler
Vercel News
Vercel News
U
Unit 42
Engineering at Meta
Engineering at Meta
aimingoo的专栏
aimingoo的专栏
MyScale Blog
MyScale Blog
Y
Y Combinator Blog
阮一峰的网络日志
阮一峰的网络日志
爱范儿
爱范儿
Apple Machine Learning Research
Apple Machine Learning Research
博客园_首页
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
B
Blog RSS Feed
N
Netflix TechBlog - Medium
GbyAI
GbyAI
F
Fortinet All Blogs
MongoDB | Blog
MongoDB | Blog
大猫的无限游戏
大猫的无限游戏
C
Check Point Blog
M
MIT News - Artificial intelligence
D
Docker
IT之家
IT之家
Stack Overflow Blog
Stack Overflow Blog

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
9router: route Claude Code, Cursor, or Copilot through wh...
George Mihai · 2026-05-10 · via DEV Community

George Mihailov

I was crawling GitHub trending for AI infra projects when I came across decolua/9router. It's the most thoughtful "stretch your free tiers" project I've seen for AI coding tools — and it does a few things I hadn't seen combined before. Worth a 60-second rundown for anyone burning through Cursor/Claude/Copilot budgets faster than they'd like.

What it actually is

9router runs locally and exposes a single OpenAI-compatible endpoint at http://localhost:20128/v1. Your IDE / coding agent talks to it as if it were OpenAI. Behind the scenes, requests get routed to whichever provider you configured — GitHub Copilot OAuth, a Gemini CLI session, an Ollama instance, Cursor, Codex, Kiro, Qwen, and others.

The interesting parts aren't the routing itself (lots of OpenAI-compatible proxies exist — litellm, OpenRouter, etc.). The interesting parts are:

  • Combos — a virtual provider made of multiple Providers. Round-robin across three Copilot OAuth sessions and you get effectively "unlimited" throughput without ever tripping a single account's rate limit. Strategy options include round-robin and sticky-round-robin (the latter pins a session to a single conversation).
  • RTK (Reverse Token Killer) — a filter layer applied before tool output goes back to the LLM. Your agent runs ls, grep, git diff, tree. The output is verbose and most of it is noise. RTK strips it. Configurable compression levels — at 2 it filters tree-like output, at 3 it also compresses git diffs. The filters live in open-sse/rtk/filters/ if you want to add a new one.
  • Caveman mode — system-prompt compression. Rewrites instructions in terse style ("be concise, no markdown" etc.) to cut output tokens. Configured per-endpoint with three escalating levels.
  • Antigravity — a MITM proxy specifically for VS Code's Copilot extension. Installs a self-signed cert, intercepts Copilot traffic on port 443, routes through your providers instead of Copilot's backend. Requires sudo and cert trust — properly hairy, but it's there.

Mental model in one diagram

Your IDE / agent
        │
        ▼
   [Endpoint] ─── API key your tool sees
        │
        ▼
[Provider | Combo] ── round-robin / fallback strategy
        │
        ▼
  [Translator] ──── format adapter (Claude / Gemini / Kiro / Ollama / …)
        │
        ▼
[Real upstream API]

Enter fullscreen mode Exit fullscreen mode

Each upstream gets its own Translator that speaks OpenAI ↔ native both ways. Caveman wraps the input prompt; RTK wraps the streaming output before it gets sent back to the agent. The split between input-side compression and output-side compression is the right shape — most "save tokens" tools do one or the other.

Pointing Claude Code at it

docker run -d --name 9router \
  -p 20128:20128 \
  -v ~/.9router:/root/.9router \
  decolua/9router

Enter fullscreen mode Exit fullscreen mode

Open the dashboard at http://localhost:20128, configure your providers, generate an Endpoint key. Then:

export ANTHROPIC_BASE_URL=http://127.0.0.1:20128
export ANTHROPIC_API_KEY=9r_yourkeyhere

Enter fullscreen mode Exit fullscreen mode

Claude Code will route every call through 9router. Same trick works for Cursor, Cline, and anything else that respects ANTHROPIC_BASE_URL / OPENAI_BASE_URL. For VS Code Copilot specifically, you go through the Antigravity MITM mode instead.

Why this is interesting beyond "free tier hack"

Three things stuck out:

  1. The "Combo" abstraction is genuinely useful infrastructure, not just a free-tier dodge. If you're running an agent that fans out work in parallel, having an N-of-M endpoint with proper retry/fallback semantics is the kind of thing you'd otherwise build yourself. 9router has it as a primitive.

  2. RTK lives in the right place. Your agent doesn't need to see the full output of find . -name "*.ts" — it needs the relevant matches. Filtering in the proxy means your agent code doesn't have to know about it; same code works with or without RTK enabled. Putting the filter in user-space is what makes it composable.

  3. The Cloudflare Worker mode lets you deploy 9router edge-side and use it from any laptop/CI without exposing your local box. You configure providers locally, then the Worker fronts the same endpoints with a remote URL. Useful for shared team setups.

Caveats worth being honest about

  • Stitching together free-tier accounts touches each provider's TOS in different ways. Copilot's terms are the most relevant — read them. 9router is the tooling, not a license to abuse.
  • Free providers have variable latency and uptime. "Free tier across N accounts" is not a substitute for paid quota in production.
  • The Antigravity MITM mode is genuinely invasive. It's an option, not the default — and you should know what trusting a self-signed cert at the system level means before you do it.

Worth a look

If your monthly Claude / Cursor / Copilot bill is starting to feel less like a tool and more like a subscription you'd notice if you canceled, 9router is the closest thing I've seen to a real architectural answer rather than a "use this cheaper model" workaround. Repo: github.com/decolua/9router.


I ran into this through Lang Labs — a side project that packs any public git repo into a Claude Code skill you can drop straight into ~/.claude/skills/. Here's 9router as one of those skills if you want a one-page reference plus a downloadable .skill bundle.