=== LOCAL AI & AUTOMATION — June 4, 2026 ===

Welcome to Episode 2, where we dig into what's happening in local AI, agent tooling, and developer automation.

Let's start with the model that's been quietly killing it in the self-hosting community: Step 3.5 Flash. This is a sparse mixture-of-experts model, 196 billion parameters total, but only 11 billion active at any time. That means it runs on consumer hardware — Mac Studios, RTX cards — while delivering reasoning speeds up to 350 tokens per second. It scores 74 percent on SWE-bench Verified for coding and supports a 256K context window. Quantized GGUF versions are already out. If you're looking for the best model you can actually run at home right now, this is it.

Microsoft Build wasn't just about flashy demos. The Windows Agent Framework is a genuine play to make Windows the default OS for AI agents. It includes local AI reasoning, cloud PC execution, and multi-agent orchestration — all with policy-based security controls. That last part matters: enterprises won't adopt agents unless they can lock them down. This gives them the key.

OpenAI upgraded Codex — and it's not just a code completer anymore. The new "Sites" feature lets Codex generate interactive, hosted websites, dashboards, and apps directly from natural language. Type what you want, it builds and deploys it. Think of it as a workspace builder, not a coding assistant. It's rolling out to Business and Enterprise plans now.

Anthropic quietly launched an official Claude Code plugin that scans your entire project and recommends MCP servers, hooks, subagents, and automations. Most users don't even know it exists yet, but it essentially bootstraps your AI dev environment from scratch. If you've been using vanilla Claude Code, you're leaving capability on the table.

And the glue holding this all together? MCP — the Model Context Protocol — now has 97 million monthly downloads and over ten thousand public servers. Donated by Anthropic to the Linux Foundation, it's become the universal connector between models and tools. Every major agent framework now speaks MCP.

On the automation side, Make.com just rolled out MCP tools for its AI agents, letting workflows tap into external tool servers without custom integration. Mastra continues to gain traction as the TypeScript-native agent framework for Vercel deployments, and LangGraph remains the enterprise Python standard with SOC 2 compliance and human-in-the-loop support.

What to watch: DeepSeek V4, Qwen 3.5, Mistral-Small-4, and Gemma-4 are all expected this year. The race to make frontier AI run on a laptop is heating up fast.

That's your local AI and automation update for June 4th. Links and audio in the show notes. See you tomorrow.

=== END ===