The Agent Stack
The Agent Stack
Six layers run the whole company: the orchestrator, the back office, the model lanes, the worker pool, the tools, and the metal underneath. Here is each one, why it is there, and the cheaper or simpler swap if that layer is not for you.
The Conductor and every department head are Claude Code running headless, spawned into tmux and committing like a senior engineer. It signs in with OAuth on a subscription plan, so there is no API key and no per-token meter. The subscription is the payroll.
You never look at this dashboard. It runs behind the scenes: org hierarchy, tickets, heartbeats, budgets. Agents wake on a heartbeat, pull work, and report back. The cockpit you actually watch just reads its board over the API.
These lanes belong to the workers, not the orchestrator. Heavy work runs GLM-5.2, code runs GLM-5.1, light work runs a flash model, and Codex on the ChatGPT plan is the second lane. Two subscriptions and one flat key beats a per-token meter you cannot predict.
Nous Research's agent engine, living on the board as one employee with its own chat, cron, and MCP. The gateway binds to loopback behind Tailscale, so the whole worker pool runs its own model lanes without a single port open to the world.
Agents are only as useful as what they can touch. MCP plugs real software into every layer of the stack: CRM, payments, docs, content, DMs, call transcripts, the terminal itself. One config file, and every agent in the org inherits the tools.
A Mac mini in the room if you want sovereignty: Docker sandboxes every worker, and the OS answers only on your tailnet. Railway if you want zero ops: two services, a volume, auto-deploy on push. Either way, Tailscale is the only front door.
Two subscriptions carry most of it. The infrastructure is the cheap part, which surprises people.
| Layer | What I run | Rough cost |
|---|---|---|
| Orchestrator | Claude subscription, OAuth not API | flat monthly |
| Model lanes | GLM via Ollama Cloud + Codex on ChatGPT | two flat plans |
| Back office | Paperclip, self-hosted | free |
| Worker pool | Hermes, self-hosted | free |
| Tools | MCP servers, self-hosted | free |
| Metal | Railway, or a Mac mini you already own | ~$15/mo, or $0 |
No per-token meter anywhere in the stack. That is the entire point of the subscription lanes.
Then the seed file becomes your company: your departments, your agents, your funnel. That is the whole trick, and it is one file.
The OS from the post is open source. Fork it, run it locally, deploy it, make it yours. No email required, no gate.
Open the repo on GitHub →I write when I ship something real: what I built, what broke, and the configs behind it. No drip sequence, no fluff, unsubscribe whenever.
The repo below is free and public either way. The list is just where the next build goes first.
Cohort 1 sold out and is already running. Cohort 2 opens to this list first, before it goes public. Dates and pricing land in your inbox ahead of anyone else.
Want it built alongside you instead? The 1-on-1 build is open →
The same operating system in this post, running as a real product: the graph, the agents, the board, the money.
Open the demo →