Working in several companies, there is a security gap with Claude. Developers are not using devcontainers because they break many things e.g. MCP, OAuth and --chrome. So they hand over everything to the AI. And the AI removes files, reads emails and does whatever it thinks is accomplishing the task at hand.

With 45 years of development experience, what can I change?

I wanted to create all my ideas, and for that to go full --yolo in a secure way, so I created human as a firewall for devcontainers that keeps MCP and OAuth flows working. And devcontainers that are easy to setup and just work. You can get some of it with other tools, but it's a pain to install them all and make them work individually.

This evolved into all I needed for development: Context via Notion, Sentry and Linear/JIRA integration, a control plane that shows running Claude instances, assigning tickets to Claude instances, all the prompts you need for development tuned to use human as a tool and much more.

https://gethuman.sh - Open Source, MIT, Free.

This is in it's early stages, works for my use cases, gas town light.