Building software · GPT

Set up an isolated machine where an agent builds and deploys an app

Idea + starter prompt, not run by us yet

For: Developers who want an agent to build software without touching their main computer

The author set up Hermes, an agent running on GPT's Codex, on a spare second-hand PC with no inbound internet. The box held Coolify, Forgejo, a self-hosted Firecrawl and Tailscale. A single prompt asked for a SvelteKit calorie tracker. The agent made the repo, wrote code and tests, fixed failing CI and deployed it with its own Postgres.

Steps

1. Use a spare machine with nothing important on it. A sandbox limits damage but does not stop the agent wiping that box.
2. Block inbound internet access and reach it through a private network such as Tailscale.
3. Install a self-hosted git server and deploy tool, so the agent has a repo and a place to deploy.
4. Give the agent only the keys it needs, and expect it could misuse them.
5. Try one small app first, then use the app yourself to find bugs the tests missed.

Starter prompt

Trial first: build a tiny hello-world page, [trial-app], through the whole pipeline before the real app. Deploy it only to the test environment on this machine.

You are working on a spare, isolated machine. After the trial, build a small web app: [describe the app, e.g. a calorie tracker]. First write a plan: tech stack, repo structure, tests and how it will deploy. Show me the plan and wait.

Create the repo, write the code and tests, run the tests and fix failing checks in the pipeline. Stop and wait for my yes before each deploy, and before you use any key or token beyond the ones I set up for this project. A green pipeline is not permission to deploy. Tell me if you need another key. Keep a short log of what you did and why.

Original post: blog.jakesaunders.dev · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.