Vibecoding: a beginner's guide

9router: The Vibecoder's Wet Dream for AI Coding
1779614849509.png

If you're fully bought into the vibecoding lifestyle—living in Cursor, letting Claude Code drive your terminal, or having Cline architect your entire backend while you just guide the vibes—you already know the two massive buzzkills that ruin the flow state:

1. The "429 Too Many Requests" Wall: Hitting rate limits right when you're in the zone.
2. The Token Tax: Going broke because your agent decided to cat a 5,000-line minified JSON file or git diff your entire node_modules folder, torching your context window and your wallet.

Enter 9router (https://github.com/decolua/9router). It’s an open-source local proxy that sits between your AI coding tools and your API providers, and it is absolute godsend for maintaining the vibe.

Here is why it's the ultimate vibecoder tool:
1. The RTK Token Saver (Compression Magic)
Vibecoding generates a lot of noise. When your agent runs terminal commands (ls, grep, git diff), the raw output eats massive amounts of tokens. 9router intercepts these tool results and seamlessly compresses them before they hit the LLM. You save 20–40% on input tokens, which means your context window stays
pristine for actual code, and your API bill drops drastically.
2. Caveman Mode: Shutting the LLM Up
Nothing kills the vibe faster than an LLM writing a three-paragraph essay about how excited it is to help you update a CSS class. Caveman Mode intercepts prompts and forces the LLM to use ultra-terse, hyper-technical language. You just get the code and the diffs. It saves up to 65% on output tokens and makes
your agents feel infinitely faster and less annoying.
3. Zero-Downtime Smart Routing
When your Claude Pro runs out of juice, the vibe usually stops. 9router fixes this with a smart 3-tier fallback system:
* Tier 1: Your premium keys (Claude, OpenAI).
* Tier 2: Cheap, fast high-throughput models (DeepSeek, MiniMax) when Tier 1 rate-limits.
* Tier 3: Free providers (OpenCode, Kiro) as a last resort.
It seamlessly routes the traffic. You never see an error; the agent just keeps coding.

4. Format Translation = Zero Friction
Want to use Gemini 2.0 or a weird local Ollama model in a tool that only supports OpenAI? 9router translates the API formats on the fly. Point your tool at http://localhost:20128/v1, tell it what model you want, and 9router handles the translation.

Get Back to Vibing
It runs locally on your machine and takes 10 seconds to set up:

1 npm install -g 9router
2 9router
Point your AI tool's Base URL to http://localhost:20128/v1 and you're done.

If you want to stop thinking about API quotas and just keep the code flowing, check it out: https://github.com/decolua/9router
 
Last edited:
Anthro said:
9router: The Vibecoder's Wet Dream for AI Coding
View attachment 1910310

If you're fully bought into the vibecoding lifestyle—living in Cursor, letting Claude Code drive your terminal, or having Cline architect your entire backend while you just guide the vibes—you already know the two massive buzzkills that ruin the flow state:

1. The "429 Too Many Requests" Wall: Hitting rate limits right when you're in the zone.
2. The Token Tax: Going broke because your agent decided to cat a 5,000-line minified JSON file or git diff your entire node_modules folder, torching your context window and your wallet.

Enter 9router (https://github.com/decolua/9router). It’s an open-source local proxy that sits between your AI coding tools and your API providers, and it is absolute godsend for maintaining the vibe.

Here is why it's the ultimate vibecoder tool:
1. The RTK Token Saver (Compression Magic)
Vibecoding generates a lot of noise. When your agent runs terminal commands (ls, grep, git diff), the raw output eats massive amounts of tokens. 9router intercepts these tool results and seamlessly compresses them before they hit the LLM. You save 20–40% on input tokens, which means your context window stays
pristine for actual code, and your API bill drops drastically.
2. Caveman Mode: Shutting the LLM Up
Nothing kills the vibe faster than an LLM writing a three-paragraph essay about how excited it is to help you update a CSS class. Caveman Mode intercepts prompts and forces the LLM to use ultra-terse, hyper-technical language. You just get the code and the diffs. It saves up to 65% on output tokens and makes
your agents feel infinitely faster and less annoying.
3. Zero-Downtime Smart Routing
When your Claude Pro runs out of juice, the vibe usually stops. 9router fixes this with a smart 3-tier fallback system:
* Tier 1: Your premium keys (Claude, OpenAI).
* Tier 2: Cheap, fast high-throughput models (DeepSeek, MiniMax) when Tier 1 rate-limits.
* Tier 3: Free providers (OpenCode, Kiro) as a last resort.
It seamlessly routes the traffic. You never see an error; the agent just keeps coding.

4. Format Translation = Zero Friction
Want to use Gemini 2.0 or a weird local Ollama model in a tool that only supports OpenAI? 9router translates the API formats on the fly. Point your tool at http://localhost:20128/v1, tell it what model you want, and 9router handles the translation.

Get Back to Vibing
It runs locally on your machine and takes 10 seconds to set up:

1 npm install -g 9router
2 9router
Point your AI tool's Base URL to http://localhost:20128/v1 and you're done.

If you want to stop thinking about API quotas and just keep the code flowing, check it out: https://github.com/decolua/9router
Click to expand...
Jesus so much band-aids for one thing...
 
1779642754230.png

Nothing better than free my dude, even if its an entire medicine cabinet full of band-aids DILLIGAF - that cost me ZERO
 
Anthro said:
9router: The Vibecoder's Wet Dream for AI Coding
View attachment 1910310

If you're fully bought into the vibecoding lifestyle—living in Cursor, letting Claude Code drive your terminal, or having Cline architect your entire backend while you just guide the vibes—you already know the two massive buzzkills that ruin the flow state:

1. The "429 Too Many Requests" Wall: Hitting rate limits right when you're in the zone.
2. The Token Tax: Going broke because your agent decided to cat a 5,000-line minified JSON file or git diff your entire node_modules folder, torching your context window and your wallet.

Enter 9router (https://github.com/decolua/9router). It’s an open-source local proxy that sits between your AI coding tools and your API providers, and it is absolute godsend for maintaining the vibe.

Here is why it's the ultimate vibecoder tool:
1. The RTK Token Saver (Compression Magic)
Vibecoding generates a lot of noise. When your agent runs terminal commands (ls, grep, git diff), the raw output eats massive amounts of tokens. 9router intercepts these tool results and seamlessly compresses them before they hit the LLM. You save 20–40% on input tokens, which means your context window stays
pristine for actual code, and your API bill drops drastically.
Click to expand...
Rtk and caveman both suck,Rtk especially caused task degradation.
 
Anthro said:
What do you suggest ?
Click to expand...
Niether, myself and a few guys on the discord did a deep dive into token compression and other token saving methods. While they save tokens the results they produce are less than desirable. Rtk being one of the worst offenders, it started causing massive cache misses, bad responses for complex tasks. Caveman was a bit better but after a while it started fumbling around with complex tasks Ultimately I just dropped all the compression gimmicks all together, dropped all contextual memory layers (including mine). Used harness engineering, proper task and plan management and I'd about 95% of my issues resolved themself.

But you're still at the mercy of the providers when they lobotomize the models like they did for the last two weeks.
 
So tonight I vibecoded something I'm pretty happy with using Gemini, for a real-world task (I want to create a guide for my high school son to explore various career paths in front of him). Gemini gave me an index.html linked to a .json database, bringing in photos dynamically from Unsplash (which it found for me), using Tailwind. I kinda follow the structure but needless to say I could never have hand-coded this in two hours like I've done now: https://mymapcodes.com/careers/

Any thoughts on this? Academically I didn't try and cover every possible career, but there's at least enough there now to give my son a good taste of the varied opportunities ... and that was my goal. That meant it wasn't possible doing it all at once, I went through dozens of iterations, building out new sections and improving the interface. But vibecoding meant it wasn't stressful, and Gemini really made it pretty fun. And yes, before anybody asks, that intro splash page was completely AI-generated, just with a bit of guidance from me over what type of intro I wanted.
 
I have been using Zoo Code / VS Codium extensively with Gemma 4 12B QAT, Qwen2.5 Coder 14B, Qwen3.5 9B, GPT OSS 20B, Deepseek R1 0528 Qwen3 8B, and Phi 4 Reasoning Plus. There are some bugs using the models, like it would bug out when debugging because a capable tool doesn't know how to use the tool when it does. Also notable using Cline, and I have previously experienced it with some cloud models. I assume it has to do with the way these tools communicate with the LLM API.

I build agents orchestrating docker containers. Each app is uniquely tailored, and it is all working using only locally hosted open-source tools and APIs. In addition, I have built so many connectors now that everything is seamlessly integrated.

Meh, I guess there is a use in vibe coding. Still though, I wouldn't sell anything that I have created. I have seen some soddy vibe codes apps attempting to make a break through... Ugh.

I have also created an app, built on all my historical reporting. That has turned out to be quite a cool audititing tool, but there are better enterprise tools out there. It is only that this one understands how I am thinking.
 
Fulcrum29 said:
I have been using Zoo Code / VS Codium extensively with Gemma 4 12B QAT, Qwen2.5 Coder 14B, Qwen3.5 9B, GPT OSS 20B, Deepseek R1 0528 Qwen3 8B, and Phi 4 Reasoning Plus. There are some bugs using the models, like it would bug out when debugging because a capable tool doesn't know how to use the tool when it does. Also notable using Cline, and I have previously experienced it with some cloud models. I assume it has to do with the way these tools communicate with the LLM API.

I build agents orchestrating docker containers. Each app is uniquely tailored, and it is all working using only locally hosted open-source tools and APIs. In addition, I have built so many connectors now that everything is seamlessly integrated.

Meh, I guess there is a use in vibe coding. Still though, I wouldn't sell anything that I have created. I have seen some soddy vibe codes apps attempting to make a break through... Ugh.

I have also created an app, built on all my historical reporting. That has turned out to be quite a cool audititing tool, but there are better enterprise tools out there. It is only that this one understands how I am thinking.
Click to expand...
Why not sell it? If it has value to someone, you win. And who's to say that what you find only moderately useful, somebody else doesn't think is amazing? What you've described sounds amazing, I take my hat off for all you've accomplished. Go get yourself rewarded for that effort.
 
Foxhound5366 said:
Why not sell it? If it has value to someone, you win. And who's to say that what you find only moderately useful, somebody else doesn't think is amazing? What you've described sounds amazing, I take my hat off for all you've accomplished. Go get yourself rewarded for that effort.
Click to expand...

To be honest, I am slightly pleased with my own creations. I am not a coder; my knowledge is limited to programming microcontrollers. My latest invention is quite a bit suprising... It is a legal tool.
 
Fulcrum29 said:
To be honest, I am slightly pleased with my own creations. I am not a coder; my knowledge is limited to programming microcontrollers. My latest invention is quite a bit suprising... It is a legal tool.
Click to expand...
Go get paid, you wild thing you. Why should only teenagers with clones of existing marketplace apps get enough dollars to feed their McDonalds addiction? Whether we like it or not, AI is not only here to stay but is already creating wealth for its earliest adopters: whether that's AI slop channels on YouTube, or advertising agencies purely using AI to write scripts and even design posters and social media, the money is the same as any other money you might earn.
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X