Tool news got weird this week.

YC says AI is moving into the physical world. China just held a robot MMA fight. OpenAI turned Codex into a physical desktop controller. The theme is pretty clear: AI is leaving the chat box.

Also: we are opening the Toolfolio Discord community. If you want to share tools, find early users, swap builds, or keep up with what we are adding, join here: Join the Toolfolio Discord

Here's what's inside:

→ But first, check out today’s sponsor to support our newsletter  👇

Most support teams are experimenting with AI. Few are transforming because of it. The gap is where the most consequential decisions in support are being made, and where most teams get stuck. Hear from industry leaders on moving from pilot to production.

Watch now

Y Combinator’s latest Requests for Startups says the quiet part out loud: AI is moving into the physical world.

They are looking for startups rebuilding real-world systems across education, healthcare, defense, finance, infrastructure, and work itself. In other words, the next AI wave may not look like another chatbot. It may look like tools that control factories, hospitals, schools, logistics, and robots.

Which makes this week’s weirdest story feel a little less random: China just hosted what outlets are calling the world’s first MMA-style robot fight.

It is not exactly UFC yet. But demos are moving from labs into arenas. Robots are becoming something people can watch, argue about, share, and eventually maybe pay for.

The bigger takeaway: physical-world AI is starting to feel less like a sci-fi category and more like the next startup theme.

OpenAI and Work Louder revealed Codex Micro, a small desktop device built around the way people now work with coding agents.

Think of it less like a keyboard and more like a physical command center for Codex. It has dedicated controls for launching agent sessions, moving through tasks, approving actions, and keeping AI work close at hand.

That might sound niche, but it points to a bigger shift: AI tools are starting to escape the chat box. First they became sidebars. Then terminal agents. Now OpenAI is experimenting with dedicated hardware.

If coding agents become part of everyday work, there may be room for new physical workflows too.

Moonshot’s Kimi K3 is now out, and the more interesting story is not just that it launched. It is how strong the early benchmark sheet looks.

The headline: K3 looks like a serious frontier coding and agent model. Moonshot reports 88.3 on Terminal-Bench 2.1, almost tied with GPT-5.6 Sol’s reported 88.8. It also reports 67.5 on DeepSWE, 77.8 on ProgramBench, and 81.2 dominance on FrontierSWE.

Those numbers matter because they point at real builder workflows: fixing repos, running terminal tasks, reconstructing programs, and handling longer engineering jobs with tools. K3 is not just “good at chat.” It is being positioned as a model for coding agents, long-context work, visual reasoning, and knowledge-heavy tasks.

There is one caveat worth keeping: not every score is an apples-to-apples public leaderboard result yet. Some K3 results use Kimi Code, some comparisons use other agent harnesses, and independent writeups are still calling it “worth testing” rather than a clean universal winner.

My read: Kimi K3 has moved from hype to something builders should actually benchmark in their own workflow.

Claude had a very workflow-heavy week. Anthropic shipped a security review plugin for Claude Code, new demos showed Claude moving closer to iOS simulator workflows, and Claude Cowork added a new way to teach Claude a repeatable skill.

The Cowork feature is the most interesting one for everyday users: you record your screen while doing a task, talk through it as you go, and Claude turns that walkthrough into a skill it can run again later.

That is a much more practical version of “AI agents.” Instead of writing a perfect prompt every time, you can show the model how you work once and turn that into reusable workflow memory.

Cursor introduced Cursor Router: instead of making users choose between every model manually, the editor can route a request to the model that fits the task.

This matters because the model menu has become overwhelming. Claude, GPT, Gemini, Grok, Composer, small fast models, expensive reasoning models, code specialists, generalists. Most people do not want to think about that every time they ask for a refactor or a simple task.

Learn more about Sponsoring Toolfolio here —> Toolfolio Boosts

Quicklinks:

Find the Latest Addition to Toolfolio here —> Newest Additions
Find our Dedicated Libraries of Tools here —> Libraries
20+ Coupons can be found on Toolfolio today —> View Coupon Codes

That’s all for this edition. Hopefully you found something new to spark ideas and make your workflow smoother.

Know a friend who’d love this? Forward it their way.

Got a tool worth featuring? Reply to this email

Catch you in the next drop,
— Dudu from Toolfolio