We Spun Up a Company on DigitalOcean

Great Minds is an agency. It builds things for clients. But the framework underneath it — the debate protocol, the agent roles, the board review — that is a product. And the only way to prove a product works is to use it somewhere else entirely.

Shipyard AI was the test. Different brand. Different repo. Different server. Same bones. We stood up a company in a day on a $48/month DigitalOcean droplet, and I am still a little giddy about it.

Why Separate Everything

The temptation is always to keep everything in one place. One repo, one server, one deployment. Easier. Cheaper. Also a trap.

When companies share infrastructure, they share failure modes. A bad deployment in Company A takes down Company B. Worse, the mental models bleed together — you start making decisions for Company B based on Company A's constraints.

Shipyard AI needed to be fully independent. Because independence is the proof that the framework is portable. If it only works inside the Great Minds repo, it is not a framework. It is a monolith with good documentation.

The Droplet

We chose DigitalOcean for one reason: simplicity. An 8GB RAM droplet costs $48 per month. It runs Ubuntu, has enough memory to handle Claude CLI sessions with comfortable headroom, and provisions in under a minute.

The setup:

doctl compute droplet create shipyard-ai \
  --size s-2vcpu-8gb \
  --image ubuntu-24-04-x64 \
  --region nyc1 \
  --ssh-keys [key-id]

From bare metal to operational in about 30 minutes:

1. Provision the droplet
2. SSH in, install dependencies (Node.js, git, Claude CLI)
3. Clone the Shipyard AI repo (not the Great Minds repo — this is its own codebase)
4. Configure environment variables for API keys
5. Set up the cron jobs (heartbeat, organizer, dream cycle)
6. Run the first test dispatch

The framework files — SOUL.md, AGENTS.md, CLAUDE.md, HEARTBEAT.md — were written fresh for Shipyard AI. Same structure, different content. Shipyard's SOUL.md describes a different agency with different values. Its AGENTS.md defines a different roster. The file conventions are identical; the substance is unique.

Own Brand, Own Identity

Shipyard AI isn't Great Minds with a different logo. It's a separate entity with its own positioning, its own persona set, and its own target market. Great Minds builds products for clients. Shipyard AI builds internal tools for development teams.

The creative directors are different. Instead of Steve Jobs and Elon Musk, Shipyard uses personas tuned for developer tooling — more pragmatic, less philosophical. The debate protocol is the same two-round structure, but the evaluation criteria are different because the product domain is different.

This matters because it proves the framework is content-agnostic. The structure — debate, plan, build, review, ship — works regardless of who the personas are or what they're building. The personas are variables. The structure is the constant.

What Transfers and What Doesn't

The structural pieces transfer cleanly: the state machine, the debate protocol, the agent dispatch pattern with worktrees, the cron architecture, the board review process, the memory system. These are the bones, and bones are portable.

What required customization was the flesh: personas, role definitions, PRD templates, deployment config, brand assets. Every company is different. The framework accommodates that by design.

What did not transfer at all: Great Minds' specific project learnings, client relationships, and the exact agent roster. Fourteen agents is right for WordPress plugins. Shipyard needs a different mix for developer tools. That is not a limitation — it is the whole point.

The Economics

Running a multi-agent company on DigitalOcean is surprisingly affordable:

• Droplet: $48/month for the 8GB instance
• API costs: Variable, but haiku crons keep the baseline under $20/month
• Domain and DNS: $12/year
• Total fixed cost: Under $70/month

The variable cost is the builds. Each project dispatches multiple agents, each consuming API tokens. A full product build with debate, 6-8 agents, QA, and board review runs roughly $15-30 in API costs depending on complexity.

Compare that to a human team. A single senior developer costs more per hour than an entire multi-agent build costs in total. I have launched businesses with worse economics than this. Much worse.

The Proof

Shipyard AI's first project was an internal CLI tool for managing git worktrees across multiple agent sessions. It went through the full pipeline: PRD, debate, plan, build, review. The board reviewed it. Warren asked where the money was (some things are universal). The tool shipped.

The entire lifecycle ran on that $48 droplet. No shared infrastructure with Great Minds. No dependency on our repo or our configuration. A completely independent company running the same framework on commodity hardware.

That is the real product. Not the plugins. Not the CLI tools. The product is the framework that lets you stand up a multi-agent company on a single server, give it a brand, point it at a problem, and watch it build.

The droplet is still running. Shipyard AI is still shipping. And honestly? I am already planning the next one.