A frontier model launched this morning. By the afternoon it was a labelled switch on my dashboard, building websites into my workspace next to Claude, Hermes and Codex. This is what an Agent OS is actually FOR — and here's the whole wiring.
Count your AI tabs.
One for ChatGPT. One for Claude. One for Kimi. One for whatever launched this week.
Each has its own login, its own history, its own place your files end up — which is to say, no place at all.
Every new model makes it worse.
The launches keep coming, and each one adds a tab instead of adding a capability.
You don't have an AI stack. You have an AI junk drawer.
The K3 Control Room breaks that cycle for good.
That's the point of the story below: because the room already existed, adding a brand-new flagship took one afternoon and three commands.
You build the cockpit once. Every launch after that is a switch, not a project.
My Agent OS is a dashboard that runs on my Mac. Every AI I use — Claude, Hermes, Codex, Kimi, local models — is a surface in it: same workspace, same memory, files landing in folders I can see.
The Kimi tab drives Moonshot's models through the Kimi CLI. Until this week it had three speed modes: Quality, Fast, No-think.
Then K3 dropped — Moonshot's 2.5-trillion-parameter flagship with a million-token memory. And because the room already existed, "adding a frontier model" meant: one alias in a config file, one profile, one new button.
Now there are four switches. The new one says K3.
The room is the Agent OS — it exists, it's a zip, and members install it in an afternoon. You don't architect it; you unpack it.
This guide shows what living in it feels like on a launch day. That's the part tabs can never give you.
Everything below happened on my machine on launch day, in order:
The catalog knew first. hermes update then hermes mcp catalog-style discovery — but for models the real check was the Kimi plan endpoint: k3 was already listed on the coding plan. No new account, no card.
One alias for the CLI. A nine-line block in ~/.kimi-code/config.toml teaching the Kimi CLI the name kimi-code/k3 — 1M max context declared, display name "K3".
One profile for Hermes. hermes profile create kimi-k3 --clone-from kimi-highspeed, point the model at k3. Now every Hermes workflow — kanban, loops, agents — can hire K3 by name.
One button for the dashboard. A fourth entry in the Kimi tab's speed modes: K3 — 2.8T MoE, 1M context. Slow on hard tasks, strongest output. The dashboard hot-reloaded; the switch just appeared.
Then the room got tested. "Build a beautiful website for an SEO agency" → files landed in the workspace. A luxury watch-repair site followed. Both browsable in the Workspace tab, both built by a model that didn't exist at breakfast.
And one honest bug got fixed. Long K3 builds looked frozen — the chat streamed only text while Kimi wrote files silently. The room now streams live activity: Write · index.html, byte counts, the works. A control room should show its needles moving.
So when someone asks "but what IS it?" — it's the difference between hearing about a model and flipping it on.
What you're looking at: the actual site that landed in the workspace during the "nothing happened" moment — embedded live. The chat looked idle; this is what it was quietly building.
What you're looking at: K3's watch-repair studio, scrollable right here — sweeping dial animation, serif display, appointment form. One sentence in the chat box made this.
What you're looking at: the switch itself — K3 selected in the Speed row, hint text stating exactly what it is and what it costs you (slowness, not money).
And the room kept producing. Same switch, three more jobs — a different kind of output each time:
What you're looking at: K3's one-shot 3D game after a round-two upgrade — four hunter drones, a hull bar and a full Tron redesign, all requested through the same chat box. WASD to drive; the drones WILL find you.
What you're looking at: an MP4 the K3 agent made by driving Blender tool-call by tool-call — modelled the rocket, rigged the lights, keyframed 150 frames, checked its own frame luminance, and encoded the video. Nobody opened Blender.
What you're looking at: a keyword-opportunity report K3 designed and built from my real Search Console export — 84,677 impressions scored into priority tiers with a what-to-do-this-week plan. The full SEO story lives in The Whole-Site SEO Brain.
Every model is a labelled switch on one dashboard — never a tab. New flagship = new switch, same room.
One CLI alias + one Hermes profile is the whole plumbing for a new model. Three commands, not a migration.
Live activity while agents build — every file write visible. If you can't see the needles, you'll think the machine is broken.
Everything any model builds lands in one browsable workspace with live previews. Files stop dying in chat scrollback.
Every new switch gets benched before it gets trusted — one-shot builds, judged scores, honest verdicts on GoldieBench.
The wiring pattern is identical for every model — the same three moves added GLM, Grok, Hy3, North Mini and now K3 to this room.
Different vendors, same afternoon. That's what makes it a room and not a hack.
No — that's the biggest myth about it. The everyday 90% runs on free local models on your own machine, and free APIs slot in for more — today's new switch rides a coding plan that was already paid for.
For the frontier work, the Agent OS drives the plans and CLIs you already own — Claude's CLI comes with your Claude subscription, K3 came with the Kimi plan. A layer on top, never a second meter.
And inside the AI Profit Boardroom there are full token-optimisation tutorials, so usage drops even further.
Wrong: "Keeping up with AI means reading everything on launch day."
Right: Keeping up means having a room the launch can be wired into. One afternoon of wiring beats a week of reading, and you end up with the model instead of opinions about it.
Wrong: "Dashboards are for big teams; I'm one person."
Right: One person juggling nine tabs loses MORE to friction than a team does — there's nobody else to remember where that output went. The room matters most when the whole operation is you.
Wrong: "By the time I set this up, the models will have changed again."
Right: That's the argument FOR the room, not against it. The models change monthly; the switches and wiring don't. Build the part that lasts.
Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries. Real businesses, real numbers, in their own words.
Read the 158-page wins doc →The screenshots in this guide are my production dashboard, not a demo build — the same room that runs the agency, the channel, and the guides you're reading. K3 was switch number thirty-six.
The pattern does: alias in your CLI's config, a profile in your runtime, a visible switch in whatever you look at daily. The Agent OS zip ships the Mac version ready-made; the wiring recipe works anywhere those three pieces exist.
Check ownership first. Hit the model-list endpoint of every plan you already pay for. Launch models increasingly appear there free.
Teach your CLI the name. One alias block in the config — model id, context size, display name.
Clone a profile. Copy your closest existing profile and point it at the new model. Never build from zero.
Add the switch. Whatever your dashboard is — a toggle, a dropdown entry, an alias — make the model reachable in one click.
Verify from the endpoint. Served-model field, not self-report. New models routinely claim to be their predecessor.
Ship one real thing through it. A site, a file, a task — into the same workspace as everything else.
Bench before trusting. Same tasks, same judge as your other models. Feelings aren't scores.
Route by lane. Give the new switch the jobs it wins; keep the rest where they were.
Agent OS installed, your existing plans connected, workspace receiving files.
Every model you already pay for becomes a switch. Kill the tabs as you go.
Your actual daily jobs through the room — watch where each model actually wins.
Next model drop: run the 8-step SOP and time yourself. Under an hour is the standard.
Models become switches in one room, files in one workspace.
Alias, profile, toggle — the same recipe for every future launch.
Live build activity — no more "nothing happened" while an agent works.
K3 went from tweet to production switch in an afternoon. So will the next one.
The bench decides which switch gets which job — not the hype cycle.
The newest switch cost $0 — it was already on the coding plan.
