We love it when someone outside the Tines bubble takes 3B for a real spin and comes back with such a cool perspective. And this one from Adam at XDA is a great example: “I finally found a vibe-coding tool I can trust with my home lab, and the AI never sees a single API key”
Adam built a rental property tracker, a price watcher, and a home lab monitoring dashboard (servers → one dashboard → email alerts if a disk fills up or a ZFS pool degrades) in about 20 minutes of back-and-forth.
A few highlights that stood out to us Tines 3B nerds over here:
- “Not a single API key actually touches the code.” Credentials are injected by the connector at runtime, so even if a model gets prompt-injected or hallucinates, there’s nothing to exfiltrate.
- Every step is its own container, complete with its own
config.tomlandDockerfile; truly isolated, not just logically separated. - Branches map 1:1 to git branches, and Autofix/Autotune build fixable “pull requests” off to the side without ever touching what’s live.
- A custom gVisor-based sandbox means steps don’t talk to the host kernel directly; “cross-contamination between runs or users is architecturally impossible.”
- A tiny, ~4.5MB static binary tunnel for connecting to home networks. No VPN interface, no NET_ADMIN, no ports published. Just a userspace proxy that dials out to an allowlisted CIDR range you control.
- The AGENTS.md spec even tells the model how to talk to non-technical users… asking about data ownership, not “concurrency” or “exclusive writers.”
His conclusion says it all: “It’s the first workflow that’s built end-to-end to take advantage of the strengths of AI coding while also handling the weaknesses in the best way I’ve seen anyone do it.”
Totally worth the full read if you’re curious how 3B thinks about security for AI-generated automations. 