Run a self-hosted, sandboxed software factory

Tools

Codex

Overview

Building an (almost) fully self-hosted, sandboxed, agentic software factory

Workflow

The author built a self-hosted, sandboxed agentic software factory on a dedicated 2021 i7 server to isolate LLM execution from their primary homelab. Using Coolify, Forgejo, and Hermes (powered by Codex), the agent autonomously executed the full SDLC: researching stacks, writing code/tests, committing to Git, running CI, and deploying to a private HTTPS URL via DNS-01 SSL challenges. The human provided a single initial prompt and intervened only once to fix a CSRF bug after the initial deployment, keeping the agent structurally contained on sacrificial hardware.

Reported result

The agent successfully deployed a calorie tracking app from a single prompt, handling repo creation, testing, CI, and deployment without further input. It resolved a post-deployment CSRF issue after one additional prompt. The author reports the workflow worked as intended, though no quantitative metrics on time saved or cost efficiency beyond the £20 Codex subscription are provided. The system remains experimental, with the author noting it is not 'harmless' due to potential credential leaks or token burn.

Human role

The original post does not say how a person stayed involved.

Limitations

The setup requires a dedicated physical server to act as a sacrificial sandbox, as the agent can still delete repos, leak credentials, or burn inference tokens. External dependencies like OpenAI Codex, Tailscale, and Porkbun DNS are required, meaning it is not fully offline. The author notes that adding API keys for external services weakens the sandbox isolation. Future improvements like VLAN segmentation and automated backups are suggested but not implemented.

Try it

I want to replicate a self-hosted agentic software factory on a dedicated, isolated server. My context is [replaceable: e.g., building a personal CRUD app]. I have a server with [replaceable: specs, e.g., 32GB RAM] and access to [replaceable: LLM API, e.g., OpenAI Codex]. Please guide me through setting up Coolify, Forgejo, and an agent interface (like Hermes) on this box. Ensure the network is isolated from my main LAN using Tailscale and Pi-hole for DNS. Configure DNS-01 SSL challenges using [replaceable: DNS provider, e.g., Porkbun] to avoid public A records. Once the infrastructure is ready, I will provide a single prompt to build a [replaceable: app type] app. The agent must autonomously create the repo, write code/tests, run CI, and deploy. Before executing the build, perform a dry run: verify the agent can access the Git instance and CI runners without external ingress. Show me the configuration files for Coolify and the agent's skills. Wait for my explicit confirmation before sending the initial build prompt. If the agent fails, do not retry automatically; report the error and ask for human intervention. Do not expose any API keys in the logs.

More cases · Software & development

• Update a code base with Copilot chat and agent mode
• Check code against user intent with review agents and tests
• Build a budgeting app from a single prompt
• Build a mental models framework with a multi-agent dev team
• Build a instant-opening markdown viewer for macOS