Vibeserver is an mostly occasionally HTTP compliant webserver that will answer all requests by invoking an LLM which generates the response completely on-the-fly.
Only very basic scaffolding exists to slightly increase the chances of a parseable response – the rest is all left to the LLM's imagination based on request parameters (URL, hostname, headers including the HTTP referer etc)!
For a limited time, a demo instance is available on https://vibes.lxgr.net/. It's single-threaded and uses a free hosted LLM backend, so please go easy on it!
It's protected by HTTP basic authentication to provide some protection against crawlers etc., but importantly to avoid confusing. The server will literally generate anything, including liability disclaimers or a lack thereof, a permissive robots.txt file, a legal imprint etc.
By entering the following credentials and visiting vibes.lxgr.net, or following any of the links below, you do so under the understanding that anything hosted there exists for entertainment purposes only and is automatically generated by a process outside of my control. No contractual or legal agreements expressed there are binding in the real world. All data you provide will be routed to an external LLM inference provider and may be retained or published by that third party.
- Username:
vibesonly - Password:
iconsent
- Install uv or otherwise install the Python packages llm and your preferred model adapter. I've successfully tested llm-mlx locally on a Macbook Pro and llm-openrouter as an API-based service.
- For API-based models, you'll need to set up an API key
- Local models have to be instlled per your plugin's instructions. For
mlx-llm, runningllm install llm-mlxand thenllm mlx download-model <modelname>should do the trick. Non-thinking models work best for acceptable latency.
- Configure your desired port number as
PORTand model name asMODEL_NAME.
➜ ~ curl http://localhost:3000/api/myip
{
"ip": "2<redacted>:1"
}%
https://vibes.lxgr.net/blog/2025/05/go-got-it-wrong-why-null-strings-are-essential.html
https://vibes.lxgr.net/hackernews/news?p=2
Leave a message!
https://vibes.lxgr.net/guestbook.php?snow=true&animate=true&color=pink
It knows GLSL!
https://vibes.lxgr.net/webgl/shader-demos/simple-rotating-triangle.html
https://vibes.lxgr.net/sigbovik/submissions2025
https://vibes.lxgr.net/scp-wiki/scp-3125
- Image generation
- Threads? (but it'll just eat through my token budget faster)
- Persistence? (probably not)
- Make it self-hosting (i.e. replace the Python script with an LLM prompt for it, then execute the result), or at least back as many methods as possible by https://github.com/awwaiid/gremllm/
The output of this web server is inherently unpredictable. It might generate things you do not agree with or want to have hosted on your website.
It will also serve all incoming requests, including those for robots.txt, and it might happily invite crawlers in that could then quickly churn through a prepaid LLM API key's budget, or rack up high costs on a billed one.
Access control is accordingly advisable for several resons.
See also [license.txt].