This week the same name showed up in my feed for the fifth time: Omarchy — DHH's opinionated Arch + Hyprland distribution. The numbers are absurd for a project that started as a dotfiles repo in mid-2025: roughly 40,000 GitHub stars (39,846 when I checked on Sept 10), an MIT license, a foundation with multi-million-dollar backing, DigitalOcean committing $1M a year, and an install that officially takes 35 seconds.
But the feature that actually matters isn't the eye candy. It's the first-boot screen.
Omarchy is the first "agentic OS" to go mainstream: on day one it asks you to pick a default AI agent — Claude Code, Codex, OpenCode, Copilot CLI, Crush, Grok, Hermes, and more. I counted the picker on the official site: 13 agents, preinstalled and wired into the system. Crash notifications get routed to your agent for diagnosis. The OS literally invites the machine to fix itself.
I think this is genuinely the future. I also think there's a bill hiding under it that nobody has priced yet.
The evening of day one
Here's how day one actually goes for the target audience. You install Omarchy in under a minute. You pick your agent. You're productive — and then, being the kind of person who installs an agentic OS, you start connecting MCP servers to it. Git ops. Containers. A database client. A web fetcher. That's the whole point, right?
And here's the mechanism almost nobody talks about: every MCP server you connect ships its entire tool catalog — names, descriptions, nested JSON Schema for every parameter — into your system prompt. Not once. Every single turn of every conversation.
Your context window starts each session paying a fixed cover charge before any real work happens.
I measured a realistic "day one" stack
To put a number on it, I built a realistic day-one developer stack — the kind a new Omarchy user would plausibly wire up in their first evening:
- 6 servers / 32 tools: git operations, container control, package management, system probes, editor RPC, HTTP debugging
- Rendered as real MCP tool definitions and measured with tiktoken (cl100k_base), the same tokenizer family most of these agents use for billing
The receipt:
| Stage | Tokens |
|---|---|
| Raw JSON manifests (what the model receives) | 5,359 |
| After schema simplification | 4,655 (−13.1%) |
| After simplification + compact TOON-style packing | 3,446 (−35.7%) |
A third of the day-one tax, gone, with zero capability loss — every tool still present, just not dressed in hundreds of tokens of redundant JSON ceremony.
Now scale the habit. An agentic OS that onboards you to 13 agents is also a machine that makes adding server #7, #8, #9 frictionless. The tax is linear in tools and billed per turn. It's the only cost in AI that recurs even when you're just thinking.
Two wrong turns I took first
Before measuring, I did what everyone does: blamed the model. The answers were getting lazy and generic, so I assumed the provider had quietly downgraded something. I even switched models for a week — same fog.
Then I blamed context drift and started nuking sessions every morning, keeping summaries. That made it worse, because a fresh session still pays the full fixed fee — I'd just multiplied how many times per day I paid it.
The actual culprit was sitting in my own config: a tool manifest dressed in nested schema ceremony. Most of the weight isn't even the descriptions you can read — it's the braces, the type nesting, the repeated parameter boilerplate around every tool.
What I use now
After fixing my own config I went looking and found a small open-source project called mcptoon (200-ish stars — genuinely tiny, which is why you've probably never seen it). It runs exactly this pipeline: simplify every tool definition, then repack what's left into a denser format. That's what produced the numbers above; I measured with its real pipeline, not a hand-rolled approximation.
Honest limits: if you run one or two lightweight servers, you don't need any of this — don't optimize a bill you're not paying. The math only starts to matter around ten-plus tools, and it gets loud fast after that.
The part I keep thinking about
Omarchy is not the problem. If anything, DHH's team did the hard part: they made the agentic OS approachable. The unfinished business is that the MCP ecosystem still bills you in tokens for furniture you installed once and never look at.
An OS that ships 13 agents on day one deserves a manifest you've actually audited.
So, honest question for anyone running an agentic setup — Omarchy or otherwise: how many MCP servers does your daily-driver agent carry, and do you know what its manifest costs you per turn?
I didn't know either, until I counted.
Top comments (0)