Claude Code Can Now See Your Screen, Click Your UI, and Test What It Built

🧠 LAUNCH

Claude Code Can Now See Your Screen, Click Your UI, and Test What It Built.

Claude Code gains computer use in research preview — the CLI agent can now open your apps, navigate interfaces, click through UIs, and visually verify what it just built. This closes the loop that every coding agent has been missing: instead of generating code and hoping it renders correctly, Claude can look at the result and iterate. Available now on Pro and Max plans via /mcp. If you've been waiting for the moment coding agents stop being blind, this is it. (40,041 likes | 3,024 RTs) Read more →

Claude Code Auto Mode Ships for Enterprise and API Users.

Auto mode lets Claude Code execute multi-step tasks without asking permission at every turn — think full feature implementation, not single-file edits. Enterprise teams and API users can enable it now with claude --enable-auto-mode. The combination of computer use plus auto mode means Claude Code can now autonomously write code, launch the app, visually check it, and fix what's broken — no human in the loop until it's done. (2,521 likes | 166 RTs) Read more →

Qwen3.5-9B-Uncensored is the local model everyone's downloading. A capable, unfiltered 9B model that runs on consumer hardware — researchers and developers who need unrestricted outputs for creative or red-teaming work have a new default. The open-source community is voting with downloads. (2,570 likes | 236 RTs) Read more →

Mistral unveils Voxtral TTS, the Forge platform, and drops hints about Mistral 4 in a deep technical interview with leadership. The only European lab consistently shipping frontier models is expanding into voice synthesis and developer tooling simultaneously — watch the Mistral 4 roadmap details for where they're aiming next. Read more →

🔧 TOOL

OpenAI's Codex Now Runs Inside Claude Code as a Plugin.

OpenAI Codex ships a plugin for Claude Code: /plugin marketplace add openai/codex-plugin-cc gives you /codex:review, /codex:adversarial-review, and /codex:rescue — all inside your existing Claude Code workflow. Let that sink in: OpenAI built a tool that runs inside Anthropic's CLI. The AI tools aren't siloed anymore; they're becoming composable layers. This is what interoperability looks like when companies decide the workspace matters more than the walled garden. (2,764 likes | 203 RTs) Read more →

ARC (Agent Remote Control) gives you a browser-based dashboard to monitor and control any local AI agent — start a task on your workstation, then check in from your phone or tablet. Open source with native Nous Hermes integration. Solves the "agent running on my desktop while I'm at lunch" problem. (83 likes | 5 RTs) Read more →

Unusual Whales MCP Server streams live options flow, equities data, and prediction markets to any AI agent. Real-time financial data via MCP is exactly the kind of integration that turns AI assistants from toy to tool for anyone doing market analysis. (321 likes | 19 RTs) Read more → For more on building MCP integrations: Create an MCP Server

Hugging Face turns arXiv into agent-readable research infrastructure. hf papers [search, read] gives any AI agent structured access to academic papers — search, retrieve, and parse without scraping. Essential plumbing for research-oriented agent workflows. (783 likes | 95 RTs) Read more →

📝 TECHNIQUE

An Anthropic Engineer's Guide to the Claude Code Features You're Not Using.

An Anthropic engineer curates their most-used hidden features in Claude Code — and with 20K likes, the thread clearly hit a nerve. Most users are running maybe 30% of what's available. With computer use, auto mode, and plugin support all landing this week, the gap between "using Claude Code" and "using Claude Code well" just got wider. Go through the list and close it. (19,827 likes | 2,088 RTs) Read more →

Perplexity shares production lessons from running voice agents at scale with OpenAI's Realtime API — context management, audio pipeline architecture, and turn-taking patterns. This isn't demo-ware; it's operational wisdom from a team handling real traffic. If you're building voice, study their context windowing approach. (260 likes | 24 RTs) Read more →

200K+ vibe-coded projects get created every day, and almost none find users. The distribution problem is real: building is now trivial, but getting noticed isn't. This thread covers 7 concrete strategies including MCP servers as distribution channels and programmatic SEO — pick one and ship it this week. (492 likes | 34 RTs) Read more →

🔬 RESEARCH

Two RCTs confirm: AI as tutor helps learning, AI as shortcut kills it. Same technology, opposite outcomes — students with unrestricted AI access accidentally bypassed learning entirely, while students using AI specifically prompted as a tutor showed measurable improvement. The difference is 100% prompt design. If you're building educational AI, the "how" matters infinitely more than the "whether." (363 likes | 55 RTs) Read more →

How AI is reshaping mathematical reasoning and human cognition — a serious academic treatment, not hype. The paper builds frameworks for understanding what shifts when machines handle the computation layer, and what that means for how humans think about proof, intuition, and discovery. (190 likes | 74 RTs) Read more →

💡 INSIGHT

Anthropic acknowledges Claude Code users are hitting limits faster than expected. With computer use, auto mode, and plugins all dropping in the same week, demand is spiking hard. If your workflow depends on Claude Code, plan for throttling — schedule heavy sessions off-peak and batch your requests. The team says they're actively investigating. (8,237 likes | 405 RTs) Read more →

Redpoint data: 46% of enterprise CIOs would replace incumbents with AI-native startups. That number is far higher than anyone expected. The enterprise replacement window isn't theoretical — nearly half of decision-makers are actively open to ripping out existing vendors if an AI-native alternative exists. If you're building in enterprise SaaS, the timing has never been better. (416 likes | 24 RTs) Read more →

The engineering career ladder is missing its middle rungs. As AI handles the mid-level tasks that used to be the training ground — feature implementation, debugging, code review — junior engineers lose the path to senior, and senior engineers face an entirely different skill profile. If you manage engineers, your growth frameworks probably need rewriting. (68 likes | 16 RTs) Read more →

🏗️ BUILD

The definitive local model guide maps your hardware to exact model recommendations — 64GB gets you Qwen3-coder for Claude Code integration, 32GB runs Qwen3.5 comfortably, 16GB handles capable chat models, and even 8GB has solid options. Bookmark this as your local model shopping list. (3,131 likes | 228 RTs) Read more →

Sebastian Raschka's "Build a Reasoning Model From Scratch" is now complete and available in early access. Given his track record with Build an LLM From Scratch, this is likely the definitive hands-on guide to understanding how chain-of-thought and reasoning actually work under the hood. If you want to understand reasoning models at the weight level, start here. (2,425 likes | 246 RTs) Read more →

🎓 MODEL LITERACY

GUI Agents (Computer Use): Claude Code's computer use launch marks a shift from API-only coding agents to GUI agents — AI that can see screens, click buttons, and verify output visually. Traditional coding agents work through structured APIs and text: they read files, write code, and run commands. GUI agents add a pixel-level perception layer — they take screenshots, identify UI elements, and execute mouse/keyboard actions just like a human would. The key tradeoff: visual verification catches bugs that text-only agents miss (a button that renders off-screen, a modal that blocks interaction), but it's slower and more compute-intensive than structured DOM access. Use computer use when you need end-to-end visual verification — form flows, responsive layouts, cross-browser rendering. Stick to API-only mode for pure code generation where visual output doesn't matter.

⚡ QUICK LINKS

Mollick: Why financial markets will feel superhuman AI first — the domain where raw intelligence converts to value fastest. (1,544 likes) Link
Contrarian take: MCP is dying, teams are reverting to plain OAuth + APIs — not everyone is seeing production-grade results. (528 likes) Link
Learn Claude Code by doing: Interactive tutorial launches — hands-on exercises instead of documentation. (109 likes) Link
Could AI agents make free software matter again? When agents maintain open-source at near-zero cost, proprietary economics may reverse. (73 likes) Link

🎯 PICK OF THE DAY

OpenAI shipping a plugin that runs inside Anthropic's CLI isn't generosity — it's a land grab. The Codex-in-Claude-Code move is the most strategically fascinating thing that happened today. OpenAI is betting that the default AI workspace matters more than the model layer, and whoever owns the developer's terminal wins regardless of whose weights run underneath. By making Codex available inside Claude Code, OpenAI gets distribution to Anthropic's most engaged power users without building a competing CLI. Anthropic gets a richer plugin ecosystem that makes Claude Code stickier. Both sides win in the short term — but the long game is about which layer captures the lock-in. If developers start relying on /codex:review inside Claude Code, does that make them more loyal to Claude Code (the workspace) or to Codex (the capability)? History says the workspace wins — just ask every IDE plugin that outlived the editor it was built for. The AI tool wars just entered their interoperability era, and the real competition isn't model vs. model anymore. It's terminal vs. terminal. Read more →

Until next time ✌️