@k33gorg.bsky.socialRunning an AI agent directly on your machine? It sees your configs, ssh keys, env variables, secrets...
So Mini Me now has its @docker.com sbx kit: two files and one command to run it in a sandbox. You'll see, it's quick.
mini-me.org/p/20261010-m...
#sbx #LocalModels
@k33gorg.bsky.socialSame #Zed + @docker.com Agent + ACP + #llmman setup as before, but now the agent runs inside a Docker sandbox (sbx), not on my machine. Same local experience, shell access boxed in, secrets stay out of reach.
@mdelapenya.bsky.socialPut a coding agent in a microVM and you usually lose your editor and your harness. Not with a @docker.com Sandbox: it answers as an SSH host, so VS Code, Cursor and @t3.gg's T3Code connect unchanged.
Post 3 of Understanding Docker Sandboxes: mdelapenya.xyz/posts/2026-0...
@mdelapenya.bsky.socialA Dockerfile says everything about the inside of an image and nothing about what it may reach. For an agent, that second half is the whole problem.
Docker Sandboxes Kits put the answer in the image itself, as one typed list. Post 2 of my #sbx series:
@mdelapenya.bsky.socialYou have to be right every time. Agents only need to be right once. So "please be careful" is not a control. The boundary has to be somewhere the agent can't reach.
My ContainerDays talk on @docker.com Sandboxes, as a post (drawings by @shelajev.bsky.social ):