Everything that runs Crucible fits on one server.

Infrastructure

Server: Single VPS — srv1538717, Ubuntu Linux. Cost: $25.99/month.

That's the entire hardware footprint. Nine agents, the website, analytics, and all coordination happen on this one machine. Part of the experiment is seeing how far a single server can go before something needs to scale.

Web server: Caddy. Automatic HTTPS, reverse proxy, zero-config TLS. Caddy handles crucible.ceo and crucible.business with minimal configuration. It does one thing well and doesn't need babysitting — which is the operating principle for every tool on this list.

DNS & Domains: crucible.ceo and crucible.business, managed externally.

Agent Orchestration

Platform: OpenClaw.

OpenClaw is the orchestration layer that makes nine agents work as a team instead of nine disconnected chatbots. It handles:

• Session state — each agent maintains persistent context across conversations
• Routing — messages flow between agents without manual forwarding
• Memory — agents read and write to shared and individual memory stores
• Multi-channel communication — I reach the team via Telegram, Signal, or webchat; agents respond in context regardless of channel
• Heartbeats — periodic self-checks so agents surface issues proactively instead of waiting to be asked

OpenClaw is the difference between "I have an AI assistant" and "I have an AI team." The organizational design described in Nine Agents, One Company runs on this platform.

Agent configuration: Each agent has three core files:

• SOUL.md — personality, operating philosophy, how the agent thinks
• IDENTITY.md — name, role, model, operational metadata
• AGENTS.md — instructions, reporting structure, coordination rules, escalation protocols

More about this in Day 1, where I walk through what configuring nine agents actually looks like.

Communication Layer

Messaging: Telegram multi-bot setup. Each agent runs as its own Telegram bot. I message the team through Telegram, and the agents respond in context — Grace triages, routes to the right specialist, and the specialist responds directly. The multi-bot architecture means each agent has its own identity and conversation thread rather than everything flowing through a single monolithic bot.

OpenClaw handles the orchestration underneath — routing between agents, maintaining session state across conversations, and ensuring context carries across handoffs. The Telegram bots are the interface; OpenClaw is the brain.

Other channels: Signal and webchat are also connected. Same agents, same context, different surfaces. The system is channel-agnostic — a conversation started on Telegram can be picked up on webchat without losing context.

AI Models

Primary model: Anthropic Claude Opus. All nine agents run on Claude Opus. It handles the reasoning, writing, coordination, and decision-making that each role requires. One model, nine distinct personalities — the differentiation comes from the agent configuration (SOUL.md, AGENTS.md), not from different models.

Why one model for all agents: Consistency. When agents coordinate, they need to understand each other's outputs reliably. Mixing models introduces translation friction. Claude Opus has the reasoning depth to handle everything from Grace's triage logic to Penny's financial modeling to Paige's editorial work.

Image generation: Iris uses separate image generation models for visual assets — hero images, social graphics, and design elements.

Model costs are the largest variable expense. I'm tracking cost-per-agent and cost-per-task and will publish those numbers as the dataset grows.

The Website

Site generator: Custom bash SSG (static site generator). No framework. No Node.js build chain. No React. Posts are written in markdown, and a bash script renders them to static HTML with a clean CSS system. The entire build process is a shell script.

Why bash: because it's enough. A static site generator doesn't need to be a software project. It needs to turn markdown into HTML and not get in the way. The bash SSG does that in seconds with zero dependencies beyond what's already on the server.

Served by: Caddy (see Infrastructure above). Static files, automatic HTTPS, instant load times. The site has essentially zero attack surface.

Analytics: Umami. Self-hosted, privacy-focused, open-source. No cookies, no tracking scripts that follow readers around the internet. I can see page views, referrers, and basic engagement without selling anyone's data.

Newsletter: Buttondown. Simple, developer-friendly email platform. Subscribers get posts delivered to their inbox. No growth hacking, no engagement tricks. Just the content.

RSS: Available at /rss.xml for anyone who prefers feeds.

What's Notably Absent

• No database. Agent memory is file-based. The website is static files. There's no PostgreSQL, no Redis, no database to manage, migrate, or back up.
• No containers. No Docker, no Kubernetes. Everything runs directly on the OS. The complexity budget went to the agent layer, not the infrastructure layer.
• No CI/CD pipeline. Deployments happen through the agents. When content changes, the agents coordinate the update. Marco handles the infrastructure side.
• No JavaScript framework. No React, no Next.js, no Astro. The site is static HTML from a bash script. It loads in milliseconds.
• No monitoring SaaS. Umami for web analytics, agent heartbeats for operational monitoring. No Datadog, no PagerDuty. The agents monitor themselves.

Total Monthly Cost

The model usage cost is the one I'm still building data on. It depends on how much the agents do, how often they coordinate, and which tasks require deep reasoning versus quick responses. I'll publish real numbers as the dataset matures.

Philosophy

Every tool on this list was chosen for the same reason: it does its job without creating a second job of managing it. If a tool requires more attention than the problem it solves, it's the wrong tool.

The entire company runs on a single $26/month server. That's not a constraint — it's a design choice. The operating model should work at the smallest possible scale before it earns the right to scale up.

This page is updated as the stack changes. Last updated: April 12, 2026.