Issue #2 · Sent Oct 23, 2026 · Week of Oct 16–22, 2026
How we keep our picks honest, a coding model for your laptop, and Agent Skills gets a neutral home
This week we lead with something of our own: how ToolVerdict keeps its picks honest, and a new guide for people just starting to build. Then eight releases that matter for builders, each with a source.
How we keep our picks honest
Every week we recheck each listed tool's pricing, license, platforms and descriptions against its own official pages. This week's check covered all 67 tools (last full check Oct 21) and found one correction: Smithery is shutting down its connection service on October 31, 2026, while its free public directory of MCP servers stays up.
Every confirmed change goes on our public What changed log, with the before and after wording and the sources. Our first full description audit on Oct 19 fixed three descriptions, including one that was already out of date when we launched: Windsurf is now Devin Desktop.
Spot something wrong? Every tool has a Report an error link, and every report triggers a recheck against the tool's official source.
New: the Start here guide
New to building? Start here walks through the seven main building blocks most projects are made from: a first version, AI features, saving data, sign-in, payments, putting it online, and knowing when it breaks. Pick what you're making to see which blocks you need, with our pick for each and a plain-English glossary.
Open Start hereThis week's verdicts
Agent Skills, the format behind SKILL.md files, moves to the Linux Foundation
Anthropic donated Agent Skills to the Agentic AI Foundation, the same neutral home it gave MCP. Anthropic says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, support the standard, and you can keep using or contributing to it without joining the foundation. The foundation also took in OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox the same day. If you write skills for your coding agent, they now sit on a vendor-neutral standard.
Source · Linux FoundationOpenAI's Decisions API is out of beta: quick yes/no, pick-one and score answers
Instead of a full chat reply, you ask a question with bounded answers (the probability something is true, a choice from a list, or a score) and get a typed answer back. OpenAI says it's up to 10x faster than GPT-6 Luna through the Responses API, and gpt-6-luna is the only supported model for now. If your app routes tickets, tags content or picks a next step, this is a faster fit than a chat call.
Source · DocsRecursive releases Flint, an open coding model built to run on your laptop
Flint is a 40B-parameter mixture-of-experts model that activates about 2.5B parameters per token. The weights are on Hugging Face under Apache 2.0, and it works with Ollama, LM Studio, llama.cpp and vLLM. Recursive says it may need more guidance than the big closed models, so give it smaller, clearly described tasks. Worth a try if you want to code offline, keep your code on your machine, or skip API fees for everyday tasks.
SourceVercel open-sources durable-harness so cloud coding agents survive crashes
It wraps the harnesses you already use (Pi, OpenCode, Claude Code and Codex) and logs every model and tool call, so after a crash or timeout a new function resumes the turn without repeating finished work or paying twice for a finished model call. Claude Code and Codex support is experimental and needs a Vercel Sandbox. Useful if you run agents in the cloud instead of on a laptop that has to stay awake.
Source · DocsLovable now writes and runs a test when you change a rule in your app
Add or change a price, a limit or a permission, and Lovable adds a small test, runs the tests that cover the change before it reports the work as done, and keeps them in your project. It doesn't write tests for copy, styling or markup. Also new on Oct 20: Lovable can create a Supabase project for an app with no backend, in the organization and region you choose. The region can't be changed later, and Supabase bills your own organization.
SourceGitHub's package firewall screens what your agents install
Coding agents and CI runners install packages with no human watching, which GitHub says makes them especially exposed to install-time malware. The firewall sits between them and the registry and checks each request against your policies, such as a cooldown period and checks against the GitHub Advisory Database. It's npm-only for now, available to GitHub Enterprise Cloud customers, and free during the preview.
SourceGoogle open-sources its Elephant-Goldfish plugin: design first, then let fresh AI sessions poke holes
One long-running AI session (the elephant) helps you design a feature with no code written. Then fresh sessions with no shared context (the goldfish) check that they understand the design doc and hunt for gaps before any code is written. Google says thousands of its engineers use it daily. It's Apache 2.0 and targets Google Antigravity first, with other tools to come. If AI-written code keeps surprising you, it's a structured way to plan first.
SourceAnthropic previews Claude Duet, a personal assistant app that uses your Claude plan limits
Claude Duet is a new app (iOS, Android and claude.ai) for professionals and small-business owners. It sends a morning brief, checks in on your connected apps and works on tasks you hand it, and it asks before it sends a message, makes a payment, signs a document or deletes anything. It's a waitlisted research preview for US Pro and Max subscribers, included at no extra cost. It draws on your plan's usage limits, and background work like the morning brief and check-ins counts too, so you may hit your limits sooner.
Source · Help pageA popular Ask HN thread this week asked how to deal with shipping AI-written code you don't understand. Lovable's change helps with part of that for people who don't code: the rules that matter most, like prices, limits and permissions, now get a test that runs before Lovable says it's done. Tests don't replace understanding your app, but they can catch a change that breaks a price or permission rule before your users do.
SourceThe weekly verdict
New AI tools for builders, and which picks changed, once a week. We will email you a confirmation link first. Unsubscribe anytime.