📌 TL;DR

• Kimi K3 is here → the first open model I'd actually use. #1 in the world for frontend, near-Fable smarts at Sonnet prices.

• Claude Code artifacts go live → the web pages Claude builds can now pull your real data every time they load. Build your team a live dashboard - ad performance, cash flow, whatever - straight from a chat.

• Builder's notes → I gave claude access to my 1password, and why I ditched obsidian

Kimi K3: we finally have a Chinese Anthropic

I have never been a fan of open-source models so far. They are cheaper, sure, but never on the level of Claude or the OpenAI models.

Kimi K3 (from Moonshot AI) is the first one I'd genuinely use.

It's exceptional at coding, and especially at front-end design - the actual look and feel of apps and websites - where it seams to beat Claude Fable 5 and GPT 5.6 Sol.

here are some highlights…

K3 took #1 on the Frontend Code Arena (the leaderboard for building the actual UI of apps and sites) - a 17-place jump from Kimi's last model, topping 6 of 7 categories.

Overall it's more measured: 3rd on Artificial Analysis's Intelligence Index (a broad "how smart across everything" score), level with Opus 4.8, behind Fable 5 and GPT-5.6 Sol. 

And then you look at the price. K3 runs $3 in / $15 out per million tokens. That’s Sonnet-level pricing for near-Fable intelligence, about half the cost of Opus per task.

They claim it's just as strong at knowledge work too, but I haven't tested it, so don't take my word for it there.

But for coding and especially web/ app design work… K3 is the one for you to be using.

How to start using it:

Skip Kimi's own apps, grab the $40/month Kimi subscription, and run the model inside the tools you already use…

1. OpenCode (a free, open-source Claude Code clone) with your API key.

2. Or Claude Code itself via a CLI proxy (a small bit of software that reroutes it to Kimi).

3. Use the OpenRouter MCP to get claude to delegate tasks to K3

Two caveats: it's 2-3x slower than Claude, and until the weights land on July 27 everything routes through Chinese servers.

Claude Code can now build you a live dashboard from a chat

Artifacts are the live web pages Claude Code can publish straight out of a session to a private claude.ai URL. you ask it to build a little tool or page, and it hands you back a real link.

Well those pages can now call your connectors every time someone opens them (connectors = the links between Claude and your actual tools - Notion, Stripe, Linear, Google, that lot).

So instead of a snapshot frozen at the moment it was built, the page pulls current data every time it loads. And it can take actions too - post a message, update an issue.

This is a great way to build proper web apps for you and your team…

Think an ad-performance dashboard, a finance tracker, a cash-flow projector. All wired into live data from the tools you already use, built in a chat, no code.

A few things to get right:

• You need a Pro, Max, Team or Enterprise plan and Claude Code v2.1.209+.

• An artifact with connectors in it, can never be shared to a public link, on any plan. Team and Enterprise can share it inside the org.

• Local MCP servers don't work in the published page (those are connectors running on your own computer). Only claude.ai account connectors do.

Full write-up is in Anthropic's announcement.

Also this week...

• Notion now opens Markdown files → it's an "Open with" shortcut on desktop, that opens your markdown file into notion, alongside an “import to notion” option
  

• Fable 5 is coming back to your plan → from July 20 it's included in all Max and Team Premium plans (at 50% of limits), and Pro / Team Standard users get a one-time $100 credit instead. A nice sequel to Fable “leaving” your plan on the 12th.

💡 Builder's notes

1Password is now my API key manager - and Claude can use it too.

1Password is a password manager (it keeps all your logins and secrets behind one master password). What I've been doing this week is wiring it into my agents.

It has a CLI - a command-line version of the app - and that lets an agent like Claude Code pull secrets straight from your vault.

Sounds terrifying, I know. But here's the bit that makes it safe: every single time Claude tries to grab something, a pop-up appears on your screen. You approve it with Touch ID or your password, and only then does Claude get that one thing.

It doesn’t have blanket access to everything.

So now, instead of keeping my API keys in ENV files (plain text files where devs usually stash their keys, sitting unencrypted on your machine), they all live in 1Password.

Same trick's been handy with the team. Instead of pasting a login into Slack, I share it as a 1Password link that expires after 3 days.

The simplest & most effective version of a company second brain…

I've moved off Obsidian for a little while now, for two reasons.

1. Obsidian is built to be a personal second brain - you're dumping in sources from all over the internet, meeting notes, and more. That's not something you want synced to your whole company.

2. It all lives locally on your computer. So the moment you set up agents in the cloud, they can't reach any of it.

I've since set up Garry Tan's gbrain and it's been working well.

But that's not actually my recommendation for most people.

The simplest company brain is literally a folder called "context," full of the key, mostly-unchanging markdown files:

• your ideal customer

• all your offers and pricing

• what you do

• who you help

• how you help them.

• and so on…

Then whatever file-sharing you already use (Google Drive, OneDrive, Dropbox), everyone installs the desktop app so the folder lives on their own computer.

Then everyone symlinks that context folder into their AI workspace (a symlink is a shortcut that makes the folder show up inside your AI's setup, so it reads the files where they already live).

One source of truth, wired into every single person's AI.

And I mean this one literally: having run both setups, the context folder gets you 80% of the results for 20% of the work.

Getting to 100% with something like gbrain takes another 80% of work for maybe an extra 20% of output.

🧰 Tools to try

• yoinks → terminal app that downloads video from X, YouTube, Instagram, TikTok and 1,800 other sites.

• OpenKnowledge → a local-first markdown editor and knowledge base built for humans and agents.

• Cuelume → cool sounds for your next project

• kapso → gives your agent its own official WhatsApp number.

🥣 Brain food

• The Talk Sock → a noise-cancelling mask for talking to Claude (and heres a review)

• Theo's full Kimi K3 review → the 40-minute deep dive behind this week's main story.

• Codex Micro → OpenAI's first hardware product: a $230 desk keypad for steering coding agents.

• DoorDash's cli → DoorDash's CTO announced a command-line tool that lets your agent order food with real money: search stores, build a cart, check out.