Just landed — Kimi K2.7 is now an agent inside Agent OS
The Goldie Conjurer™ + Kimi Code.
Say one sentence. Get a finished app.
I added Kimi Code — Moonshot's Kimi K2.7 — to my Agent OS this week. It's not a chatbot that hands you broken snippets. You describe what you want in plain English, and it writes the entire working app, checks its own work, and drops it in your workspace. Everything in this reel was built from a single sentence:
▶ A reel of apps Kimi K2.7 built — each from one sentence. Tap any card below to open the live build.
✦
Real, finished apps from one prompt each. Click any to play it.
These are live, self-contained apps Kimi K2.7 wrote — a desktop OS, a drum machine, a playable game, generative art. No snippets, no scaffolding to finish yourself. Tap a card to open the real thing in your browser.
KimiOS — a desktop OS
Built from: "a miniature desktop operating system in a browser tab"
Beat Lab — a drum machine
Built from: "a step-sequencer + drum/synth machine"
Neon Breakout — a playable game
Built from: "a fully playable neon Breakout game"
Particle title
Built from: "particles that assemble into 'KIMI K2.7'"
Flow-field art
Built from: "generative flow-field particle art"
Metaball fluid
Built from: "an interactive neon fluid simulation"
✦
Kimi K2.7. An agent that ships whole apps.
Kimi K2.7 ("K2.7 Code") is Moonshot's coding agent — 262K context, built for long, real coding jobs. I wired it into Agent OS as a first-class agent, the same way Claude, Hermes and Codex live there. Here's why it's different:
1.It writes the whole thing — not a snippet
Most AI gives you a chunk of code to wire up yourself. Kimi writes the complete, self-contained file — UI, logic, the lot — and saves it to disk. The desktop OS above is one 39 KB file.
2.It checks its own work
Before it hands the build over, Kimi actually runs it — it loaded KimiOS in a headless browser and ran a syntax check, then reported what rendered. An agent that QAs itself.
3.It lives in your OS — chat, workspace, previews
You talk to Kimi in the Agent OS chat (with memory across turns). Everything it builds lands in your Kimi workspace, and clicking any file renders it live.
4.One plan, build all day
It runs on your Kimi coding plan — one flat rate, no per-token meter. Describe app after app; the cost doesn't move.
Thinking it?"AI can only do snippets, not real apps."
The reel above is a working desktop OS, a drum machine with live audio, and a playable game.
Each from one sentence. Open one and use it.
✦
The benchmarks. K2.7 Code, measured.
Kimi K2.7 Code is a 1-trillion-parameter mixture-of-experts model — only 32B active per token — with a 256K-token context, built specifically for long, agentic coding jobs. Straight talk on the numbers: these are Moonshot's own published benchmarks, shown as the jump from the previous K2.6. Independent SWE-bench-style scores aren't out yet — the day they are, I'll add them here.
1Ttotal params · MoE
32Bactive per token
256Kcontext window
~30%fewer thinking tokens vs K2.6
Coding + agentic — the jump from K2.6 → K2.7 Code
Kimi Code Bench v2 · coding50.9 → 62.0
Program Bench · whole programs48.3 → 53.6
MLS Bench Lite · Python / Rust / Go26.7 → 35.1
Kimi Claw 24/7 · long-horizon agentic42.9 → 46.9
MCP Atlas · tool use69.4 → 76.0
MCP Mark Verified · tool use72.8 → 81.1
The one head-to-head so far — tool use
On MCPMark (an agentic tool-use benchmark), K2.7 Code scores 81.1% — ahead of Claude Opus 4.8 at 76.4%. Tool use is exactly what matters for an agent that writes a whole app and then checks its own work — which is the whole point above.
Kimi K2.7 Code81.1%
Claude Opus 4.876.4%
"K2.7 Code uses ~30% fewer thinking tokens than K2.6 while scoring higher across Moonshot's coding benchmarks."
— Moonshot AI · Kimi K2.7 Code release, June 2026
✦
My story · why this matters
I was you. Then my ideas got built.
Before
I'd have an idea for a little tool and just... describe it to myself.
Because building it meant a dev, or learning to code, or fighting an AI that gave me broken chunks I couldn't stitch together.
I'm not a coder. So most of my ideas died in a notes app.
Then I put Kimi K2.7 in my Agent OS and just described what I wanted.
After
Now I say one sentence and a complete, working app shows up.
It's already in my workspace, running, before I've finished my coffee.
A whole desktop OS. A drum machine. A game. From sentences.
The ideas that used to die now ship.
You can have this too. Same agent. Same one-sentence builds.
✦
✦
Before you scroll on —
Commit to building one thing today.
You've seen the proof above. Real, finished apps. From sentences.
The next few minutes show exactly how it works inside the OS.
So here's the deal.
Promise yourself one thing right now — before you sleep tonight, you'll describe ONE idea to an agent and let it build the whole thing. Just one. Because the moment you watch an app you only described show up working, the gap between "I have an idea" and "it exists" closes for good.
The people sitting still are still waiting on devs. The people describing apps into existence today are the ones who look back in six months and say "that was the moment."
Be one of those people.
Commit to the transition. Commit to building today. This changes what you can make.
✦
The Goldie Conjurer™.
Four moves turn a sentence in your head into a finished, running app in your workspace.
i.The Incantation
You describe what you want in plain English — no code, no spec, no blank page. Just say it.
ii.The Build
Kimi writes the entire app itself — every line of UI and logic — into one self-contained file, and saves it to disk.
iii.The Self-Check
Before handing it over, Kimi runs and verifies its own work. You get something that actually works, not a draft to debug.
iv.The Gallery
Every build lands in your Kimi workspace inside Agent OS. Click any file → it renders live. Nothing gets lost.
✦
How most people build a tool. And how you'll conjure one.
Old way~weeks · $$$
- Have an idea, describe it, never build it
- Hire a dev or learn to code
- AI gives broken snippets you can't stitch
- Debug code you don't understand
- Most ideas die in a notes app
New way · one sentence~minutes · one plan
- Describe the app in plain English
- Kimi writes the whole working file itself
- It checks its own work before handing it over
- It lands in your OS, previews live, one click
- Every idea becomes a real, running thing
✦
Describe it, then open it.
1.Open Kimi Code → Chat
It's an agent in your Agent OS sidebar. Tell it what to build: "build me a synthwave clock", "make a habit tracker". Chat remembers your turns.
2.It writes the whole app + saves it
Kimi builds the complete file and drops it in your Kimi workspace — no copy-paste, no setup.
3.Open the Workspace — it's running
Every build sits in your workspace; click it and it renders live, right there. Open full-screen to use it.
Thinking it?"I'd need to be a coder."
You describe it in plain English. Kimi writes every line.
You never touch the code — you just open the result and use it.
✦
Run Kimi inside a full agent stack — done for you.
If you want Kimi wired into a whole system — chat, the workspace, live previews, alongside Claude, Hermes and more — it all lives inside the Agent Operating System in the AI Profit Boardroom.
- The full Agent OS zip — Kimi, Claude, OpenClaw, Hermes in one dashboard
- The Kimi setup — the agent, the workspace, the preview wiring, ready to run
- A 30-day roadmap — your first app day 1, a shelf of them by week 1
- 4 weekly coaching calls · prompts · the Obsidian memory setup
- 3,200+ founders · 38 countries · someone online 24/7
✦
Should you build one today?
If you've ever had an idea for a tool and sat on it — yes. Tonight.
It's one sentence. The first time you watch a finished app you only described show up running, the whole game changes.
The people describing apps into existence now, while it's this far ahead of normal, are the ones miles ahead when it's normal. Every build you make compounds.
✦
What you just gained.
i.
You stopped coding.
Describe it — Kimi writes the whole app.
ii.
You stopped debugging.
It checks its own work before handing it over.
iii.
You stopped waiting on devs.
A finished, running app in minutes.
iv.
You stopped losing ideas.
Every build saved + live in your workspace.
Describe the app. It builds it. It even checks its own work.
Turn every idea into a running app — by describing it.
Grab the Agent Operating System inside the AI Profit Boardroom. It turns Kimi, Claude, OpenClaw and Hermes into one system — shared memory, one dashboard, build by describing. Every new agent you add makes the whole thing more powerful.
- 3,200+ members · daily tutorials · 30-day roadmap
- The 158-page wins doc — read it here
- The Kimi setup, every prompt, the Obsidian memory setup, coaching calls where we build it together