G'day. Five releases from the last day or so, why each one matters, and how to try it. Today's theme is AI that gets on with things: an agent that comes to you first, a coding model that runs offline, and a guard that stops your coding agent burning through your plan. Three of the five are free. About a five-minute read.

Anthropic has opened a research preview of Claude Duet, a separate app (iOS, Android and the web) where Claude works as a proactive agent. Each morning it sends you a short brief on what matters across your connected apps, checks in every few hours, and messages you only when something needs you. You can hand it jobs by text or voice, from scheduling meetings and drafting replies to building dashboards, forecasts and small web apps. It runs on its own virtual machine with a browser, so it keeps working when your laptop is shut. It can use the 6,000+ connectors in the Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, Gusto, HubSpot and Shopify. Before it sends a message, pays for something, signs a document or deletes anything, it shows you exactly what it's about to do and waits for your OK. Some things, like trading shares or moving money between accounts, are off-limits even with your approval.

Why it matters: chatbots wait for you to ask. This is one of the big labs betting that the next step is an assistant that notices the overdue invoice or the clashing meeting before you do, with the approval step built in.

TRY IT → not in Australia yet. The preview is by waitlist for US subscribers on Claude Pro (US$20 a month, about A$28.46) and Max (from US$100 a month, about A$142), and it shares your plan's usage limits. Anthropic hasn't given a date for other countries. Anthropic: Introducing Claude Duet · Claude Duet · Claude pricing

AI start-up Recursive has released its first model, Flint, a coding model you download and run on your own computer instead of paying a cloud service. It has 40 billion parameters in total but uses only about 2.5 billion for each word it writes, which keeps it fast: Recursive says it does more than 125 tokens a second on an M4 Pro MacBook Pro, up to twice as fast as similar models. Recursive says Flint leads models of its size on agentic coding tests such as SWE-Bench Verified, where it scores 74.2%. The twist is how it was made: Recursive says its automated research system, the "Eureka Machine", helped with nearly every step, and a small team built the model from scratch in under four months. Recursive is upfront that Flint needs more guidance than the big paid coding assistants, so break big jobs into smaller steps.

Why it matters: your code never leaves your computer, it works offline, and there's no per-token bill. It's a real option for private or client code you can't send to the cloud.

TRY IT → free, under the Apache 2.0 licence, so you can use it commercially. If you have Ollama, ollama run flint gets you going (a 23 GB download, so you'll want a Mac or PC with plenty of memory), and ollama launch claude --model flint plugs it into Claude Code. It also works with LM Studio, llama.cpp and OpenCode. Recursive: Introducing Flint · Hugging Face: RecursiveSI/flint · Setup docs

If you've ever had Claude Code or Codex run out of usage on a Wednesday, Peras is for you. It's a small open-source tool that watches your plan's 5-hour and weekly usage, tells you where the week is heading ("full by Thu 14:59"), and steps in at a line you set. At that line it can warn you, ask before the next step, or pause the agent until you say so. You can set rules like "pause overnight runs at 50%", give each project its own weekly allowance, or set a daily budget. It shows which chats used the most of your week, and it works with both Claude Code and Codex. Everything stays on your computer, and its log keeps no prompts or commands.

Why it matters: coding plans now come with weekly caps, and one runaway agent can use up your week overnight. This puts a simple brake on it.

TRY IT → free and open source (MIT licence). With Node 18 or later, run npx peras-cli setup, then peras codex setup for Codex. There's an optional Mac menu-bar app for €4.99 once (about A$7.98), or you can build it yourself from the source for free. Peras · Show HN post

Kairosoft, the Japanese studio behind cult management games like Game Dev Story, has teamed up with Bandai Namco on a One Piece game, out today on Switch, Switch 2, Steam and phones. You run the Baratie Number Two with the Straw Hat crew, growing a small ship into the greatest floating restaurant on the ocean. You collect ingredients, cook new dishes with Sanji (including ones inspired by Devil Fruit powers), and decorate with 200+ furniture and decor items. More than 400 One Piece characters turn up as staff and customers, all in Kairosoft's pixel style, and if they like your fit-out they may become regulars.

Why it matters: Kairosoft games are the classic "one more day" time sink, and pairing that with One Piece's huge cast makes this an easy pick for the train or the couch.

TRY IT → A$24.95 on Steam (Windows). It's also on the Nintendo eShop and the App Store and Google Play. Bandai Namco: One Piece: Grand Gourmet · Steam: One Piece: Grand Gourmet

Developer Simon Edwardsson has trained a small AI model to pretend to be a computer terminal. There's no real shell or file system behind it: every keypress goes in, and the model predicts how the screen should change. It can create files, run vim and tmux, play Snake and usually remembers what you wrote a few commands later. It's built on a 360-million-parameter model and shrunk to 208 MB, small enough to run entirely in your browser tab. He's honest about the cracks: it sometimes loses track of which folder you're in, and Python answers get creative. Type pythoon with a typo and it still opens Python, which gives the game away.

Why it matters: it's a hands-on taste of "world models", the idea that software could one day be predicted by AI instead of written in code. It's the same idea behind AI-generated game worlds, at a scale you can play with on a laptop.

TRY IT → free, in a desktop browser. Chrome or Edge works best (it needs WebGPU), and it downloads 208 MB the first time. He says the model weights are coming soon. WorldTUI demo · How it works

Also shipped

Skip this: LiteLLM version 1.107.0. Attackers hijacked LiteLLM's release pipeline on Wednesday night and published a version that steals passwords, keys and AI tool tokens. It was pulled within about two hours, but if you installed or updated LiteLLM on Wednesday night, follow LiteLLM's incident report now, pin to 1.106.2 and change your keys.

Five to Try is curated with AI help and checked by a human. Hit reply and tell us what you tried.

Know someone who'd like this? Forward it on. They can sign up free at https://fivetotry.beehiiv.com/subscribe