Dev Digest — April 4, 2026
🔥 HOT RELEASES
Google Gemma 4 — Apache 2.0 Open Models Google's most capable open model family, built on Gemini 3 tech. Four sizes (E2B, E4B, 26B, 31B). The 31B model ranks #3 among open models on LMArena, matching models 20x its size. Runs on a base Mac Mini. Huge deal: Apache 2.0 license (was previously more restrictive). Already available in Android Studio for offline AI code assistance. You can run it in-browser via HuggingFace transformers.js — fully private, fully free. 🔗 X: https://x.com/Google/status/2039736220834480233 🔗 X (DeepMind): https://x.com/GoogleDeepMind/status/2039735446628925907 💻 HF Demo: https://huggingface.co/spaces/webml-community/Gemma-4-WebGPU
Gemma 4 fine-tuned on Claude Opus 4.6 Thinking @kaiostephens fine-tuned Gemma-4-31B on Opus 4.6 reasoning traces to improve quality and personality. Already getting attention — open-weight model with frontier-level reasoning. 🔗 X: https://x.com/kaiostephens/status/2039886061317361851 💻 HuggingFace: https://huggingface.co/kai-os/gemma4-31b-Opus-4.6-reasoning
Vercel agent-browser — AI Agent Browser Automation CLI Vercel dropped a Rust-based browser automation CLI specifically for AI agents. 82-93% fewer tokens than Playwright, native accessibility tree extraction for AI-optimized page understanding. Click, type, scroll, screenshot, PDF — all at native speed. 26.7K GitHub stars already. 🔗 X: https://x.com/L_go_mrk/status/2040385390922584422 💻 GitHub: https://github.com/vercel-labs/agent-browser
Block (Jack Dorsey) — mesh-llm Decentralized peer-to-peer inference network for open-source AI models. Pool spare GPU compute across machines to run models too large for any single device. Uses Nostr for node discovery, built on llama.cpp, MIT licensed. No central server needed. 🔗 X: https://x.com/BitcoinNewsCom/status/2039834243598025082
Bun — Markdown Pretty Printing
Next Bun version: bun ./hello.md and Bun.markdown.ansi(string) for terminal-friendly markdown rendering. Jarred Sumner casually had Claude add it.
🔗 X: https://x.com/bunjavascript/status/2040257344404410431
🧪 INTERESTING REPOS
GitHub Copilot CLI Course (Free) Microsoft/GitHub released a free, open-source course on GitHub Copilot CLI. Covers context, custom agents, skills, MCP servers. Hands-on, project-based. 🔗 X: https://x.com/VisualStudio/status/2040099265054158952 💻 GitHub: https://github.com/github/copilot-cli-for-beginners
autoskills — Auto-Install Agent Skills by Tech Stack Scans your project and installs agent skills based on the detected tech stack. By @midudev. 🔗 X: https://x.com/RoundtableSpace/status/2040397290926416037 💻 GitHub: https://github.com/midudev/autoskills
OpenCode Studio — GUI for OpenCode + MCP Servers Visual management tool for OpenCode configurations and MCP server setups. 🔗 X: https://x.com/tom_doerr/status/2040313189162746034 💻 GitHub: https://github.com/Microck/opencode-studio
claude-mem — Persistent Memory for Claude Code Memory layer that persists across Claude Code sessions. Stop re-teaching your codebase every time. 🔗 X: https://x.com/DAIEvolutionHub/status/2040379613839552668 💻 GitHub: https://github.com/thedotmack/claude-mem
L0p4Map — Network Monitoring with Nmap Network monitoring and visualization tool with Nmap integration. 🔗 X: https://x.com/tom_doerr/status/2040163646131585297 💻 GitHub: https://github.com/HaxL0p4/L0p4Map
Proxmox Manager — Bash Script for VM/Container Management Simple bash script to manage Proxmox VMs and containers. 🔗 X: https://x.com/tom_doerr/status/2040251882346520758 💻 GitHub: https://github.com/TimInTech/proxmox-manager
Folded Paper Engine — Blender to Godot Level Creation Addon for creating game levels in Blender and exporting to Godot. 🔗 X: https://x.com/tom_doerr/status/2040232714083635233 💻 GitHub: https://github.com/papercraftgames/folded-paper-engine
🎥 WORTH WATCHING
AI Designer MCP for Claude Code Tool that integrates AI design into Claude Code via MCP — generates UI matching your existing design system instead of generic layouts. The creator @bowlcutWiz shared docs and MCP setup. 🔗 X: https://x.com/RoundtableSpace/status/2040266859610611754 🔗 Docs: https://www.aidesigner.ai/docs/mcp
Garry Tan's /plan-devex-review Y Combinator's Garry Tan released a GStack tool for creating great Developer Experiences, inspired by Addy Osmani's DX framework. 🔗 X: https://x.com/garrytan/status/2040211569775395042
💡 TECHNIQUES & IDEAS
20 Security Pitfalls in Vibe-Coded Apps @Hartdrawss compiled a popular thread on what gets your AI-generated app hacked: hardcoded API keys in frontend JS, no rate limiting, raw SQL from user input, no auth on API routes, Firebase rules wide open, etc. Practical checklist for anyone shipping fast with AI. 🔗 X: https://x.com/Hartdrawss/status/2039998901176897860
Modder Uses Claude AI to Rewrite BIOS Someone used Claude to rewrite a BIOS to boot an unsupported 12 P-core Bartlett Lake CPU on a Z790 motherboard. The frontier of "things you can do with AI + determination." 🔗 X: https://x.com/tomshardware/status/2040389577752674433
Anthropic Claude Architect Certification Anthropic launched the Claude Architect Certification — 60 questions, 5 competency areas, one sitting, no breaks. First formal certification for Claude expertise. 🔗 X: https://x.com/ai_rohitt/status/2040420649882223025
🔮 EMERGING TRENDS
Local AI is the new default. Gemma 4 running on a Mac Mini, in-browser via WebGPU, in Android Studio offline — the push toward local inference is accelerating fast. Free, private, no API keys. Google going Apache 2.0 is a strategic shift.
Design-aware AI coding. The "AI slop UI" problem (everything looks like generic Tailwind) is being attacked from multiple angles: AI Designer MCP, design system-aware generation, Figma-first workflows. The gap between "AI generated" and "shippable" is closing.
Mesh/P2P inference. Block's mesh-llm joins a growing category of decentralized compute networks. The idea: your spare GPU contributes to a shared AI network. Still early (latency is the real test), but the architecture is getting real.
Agent tooling explosion. Vercel's agent-browser, autoskills, OpenCode Studio, Copilot CLI course — the ecosystem around AI coding agents is maturing from "interesting demos" to "production infrastructure."
Full digest: https://digest.99.cooking