How much of the Linux kernel that gets shipped in production is actually never used?
I'm wondering if release teams building/using a kernel image are often overwhelmed by the number of kernel configuration options, and end up shipping code that will never actually be used by their workload.
Are there any automatic tools out there that can fit a minimal Linux kernel to your runtime needs?
I've been building BubbleHub a local runtime and LLM hosting for agents.
1. Host specialized local models
It detects your hardware and suggests suitable base models you can run locally.
After choosing a base model you can prompt it or specialized LoRA adapters from the host and sandboxed environments.
example:
bubble prompt --text "explain this repo"
bubble prompt --text "code a calculator" --speciality coding
(Also works through a Python interface compatible with OpenAI-style usage.)
2. Multi-model scheduling + caching
BubbleHub includes a model scheduler that manages concurrent agent access to local models.
Instead of each agent independently loading/unloading models, BubbleHub keeps models warm and schedules inference requests across agents.
3. Secure sandbox ("bubbles") with dynamic permissions
Each agent runs in a sandbox.
If an agent tries to access something outside its permissions, the system pauses and requests human approval.
Approved actions are stored in the agent manifest.
The next big thing is an app store for AI agents. You’ll be able to install something like OpenClaw in BubbleHub, which automatically sets up a secure sandbox and configures any required API keys.
I'm looking for feedback from people who run local models or build agent systems: would you use BubbleHub for agent development and deployment?
If you find this project interesting, consider starring the repo or contributing.
GitHub: https://github.com/bublhub/BubbleHub
Website: https://bubblehub.ai/
Discord: https://discord.gg/skwKqSgvD2
