This week was a good cross-section of where everyone’s at. Some of us are wiring up MCPs inside real companies. Some are building agent harnesses from scratch. Some are just discovering that Webflow can quietly generate full apps. And honestly, that mix is what makes these calls good. Here’s what we covered. * * * 🔌 Rafi: GitHub + Anti-Gravity + MCP at Work Rafi got: - GitHub working - Vercel connected - GitHub connected to Anti-Gravity - And got pulled into an internal “let’s connect MCP to Fathom + Google Drive” conversation at work Which is one of those moments where you realize: > “Oh… I actually know what they’re talking about.” We talked through: - Using Google AI Studio as a shortcut builder - Dropping a PRD directly into AI Studio and letting it build a web app - Using Gmail triggers + Drive automation to process transcripts If you’re in a company environment trying to wire AI into existing workflows, this part’s very practical. * * * 📱 Android App in Two Days? Maybe. Rafi also tried building a simple Android app (a “lock social media until you pray” style blocker). Ran into emulator issues. Burned tokens. Laptop nearly cooked an egg. We talked about: - Why most “2-day app” stories are either glossed over or heavily scaffolded - Using Appwrite / React Native / Flutter-type frameworks instead of raw Android SDK - The difference between web apps and actual app store apps - When it’s worth going native vs. just shipping web-first If you’ve tried to build a “simple” app and discovered it’s not simple, you’ll relate. * * * 🌐 Amadeus: Webflow AI + Gaming Site This was one of the more interesting segments. Amadeus showed a Webflow AI-generated site for a game (Arc Raiders): - Pulling weapon data - Pulling stats - Generating layouts - Creating structured pages with almost no manual input And here’s the thing: It looked good. We talked about: - Webflow possibly wrapping Claude 3.5 - Exporting Astro code - Reverse engineering a PRD from a finished site

I quote-tweeted a thread showing how many separate repos are being built around the same problem: https://x.com/VibeCodingBot/status/2025273417369366669?s=20 How do agents use API keys without ever actually holding them? When you see this much parallel work, it usually means the need is real. I went through the repos mentioned (plus a few others) and mapped out: * Options for local OpenClaw setups * Running remote OpenClaw instances safely on a VM * Ways to avoid transferring raw secrets * Practical patterns for scoped and encrypted access If you want the breakdown, download the attachment.

Let's get to know each other! Comment below sharing where you are in the world, a photo of your workspace, and something you like to do for fun. 😊

This week’s call was a good mix of beginner questions, deep agent architecture, and some real “where is this all going?” conversations. We had five of us on: Wes, Aty, Shawn, Chris, and Gary — and the spectrum of experience in the room actually made the discussion better. Here’s what’s in the video: 🧭 “I’m Not a Developer. Where Do I Start?” Gary’s question was simple and honest: I’ve done some HTML and CSS… but with all this B-Mad, Claude Code, OpenClaw stuff — where do I even start? He’s running VS Code on a Raspberry Pi (which is awesome, by the way), trying to understand the stack without breaking his main machine. We talked about: - Not needing to become a “developer” in the old sense - Starting with outcome definition instead of tools - Keeping early builds simple (MVP mindset) - Avoiding the trap of over-architecting too soon If you’ve felt overwhelmed by: - context windows - local models - agent frameworks - “greenfield vs brownfield” talk You’ll relate to this part. 🧠 Sonnet 4.6, Codex 5.3, and the Shift in Model Power We got into the recent updates: - Sonnet 4.6 improvements - 1M context window options - Codex 5.3 becoming very test-driven - Models increasingly self-checking and structuring output There was a really interesting comparison between Claude and Codex: - Claude tends to “get it working” - Codex tends to enforce tests and longer-term structure That difference matters once your projects get big. 🏗 Chris: Building an OpenClaw Alternative (Local Model Focus) Chris shared that he’s been building his own agent framework — designed to eventually run well on local LLMs. He’s intentionally “skating where the puck is going.” Key themes: - Preparing for local models to get strong enough - Adding guardrails around smaller models - Running into scaling problems as projects grow - The importance of test coverage before things get out of control If you’re building something serious, this part is worth watching.

Anthropic released Sonnet 4.6 today. Here's what changed and why it's worth paying attention to. The biggest jump: Novel problem-solving ARC-AGI-2 measures how well a model can reason through problems it hasn't seen before — generalization, not memorization. - Sonnet 4.5: 13.6% - Sonnet 4.6: 58.3% - Increase: +44.7 percentage points That's the largest single-generation improvement in the table by a wide margin. Agentic benchmarks The benchmarks most relevant to tool use and automation all improved significantly: - Agentic search (BrowseComp): 43.9% → 74.7% (+30.8pp) - Scaled tool use (MCP-Atlas): 43.8% → 61.3% (+17.5pp) - Agentic computer use: 61.4% → 72.5% (+11.1pp) - Terminal coding: 51.0% → 59.1% (+8.1pp) Sonnet 4.6 vs Opus 4.5 Worth noting — Sonnet 4.6 now outperforms Opus 4.5 on several benchmarks: - Novel problem-solving: 58.3% vs 37.6% - Agentic search: 74.7% vs 67.8% - Agentic computer use: 72.5% vs 66.3% Sonnet is the smaller, cheaper model tier — so this shifts the cost/performance equation for anyone building agentic workflows. What this means practically If you're building with tool use, MCP integrations, or multi-step AI workflows, the MCP-Atlas and BrowseComp improvements are the ones to watch. Models that reliably use tools and follow through on multi-step tasks open up a lot of what was previously too brittle to ship.