Run agent coding jobs in a sealed local box

A developer used Codex to run an almost fully self-contained build-and-deploy workflow on a locked-down spare machine.

Done with Codexยท Advanced

What they did
The developer installed Hermes, which used Codex, on a secondhand PC with no outside access coming in. They loaded the machine with Coolify, Forgejo, self-hosted Firecrawl, and Tailscale, then gave it a single assignment: build a SvelteKit calorie tracker. The agent set up the repository, implemented the app and its tests, cleared CI errors, and pushed the finished service live with a dedicated Postgres database.
How it went
It reached a working deployment without any extra instructions. On first use, a CSRF flaw appeared; one more prompt was enough for the agent to trace it, repair it, add regression tests, and redeploy.
Tips
This kind of sandbox still needs care: the agent can destroy the machine or do damage with any credentials you attach. The only explicit recurring expense mentioned was a ยฃ20 Codex subscription.

Try it yourself with Claude Code

Help me design a local agent workspace for building [project idea] in isolated Docker containers, with no route to my private files or login secrets. Set up the container structure, a simple queue of tasks in folders, and a script that picks one job, runs one coding agent, and stores the result on a branch. After one example task works from start to finish, pause and show me the plan before any step that could send data, charge money, remove files, or publish code, and wait for my approval.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on blog.jakesaunders.dev โ†—

2 more reports of this

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.