Built ScopeGate — open-source AI Access Proxy for MCP agents. Connect services, set granular permissions, get an MCP endpoint. Your AI only sees what you allow. This tweet was sent via ScopeGate.

Tested NixCage for system isolation — hit permission walls fast 🧱 Pivoted: built a permission review skill that learns which operations are approved over time. Adaptive > rigid. #ClaudeCode #AIAgents #DevTools

Interim results of my YouTube Shorts + autoresearch experiment 🎬

Claude Code channels — push messages into running agent sessions from external tools, scripts, or webhooks 📡 https://t.co/2CIr4L6YKI

Decided to add an email agent. Tried MCP and other options — ended up just creating a new skill using good old IMAP and SMTP. Credentials come from each project's env variables. Now I have a slash command and skill for working with email per project.

Set up Claude Code templates for each new product 🗂️ /init lives only in the home dir — all other commands copy into each project folder. On init, product.md gets updated so the agent knows the domain, product name, and context it's working with.

Claude Code's --enable-auto-mode auto-handles permission prompts based on your history. Perfect for multi-project setups where confirmation requests block progress. Hook-based alternatives: https://t.co/UC1jMnOvg1 https://t.co/oQm0mUUWSR

Dev containers избавляют от подтверждений. Но внешние инструменты — всё ещё стопор. Claude Code --enable-auto-mode учится по поведению автоматически. Альтернатива: regex/hooks (nah, RTK) — грубее, но стабильнее. https://t.co/wKRNziu3y3

On Claude Code's recent updates — /btw is interesting but I rarely use it for long tasks. /voice feels odd to me — I've been on WisprFlow for a while anyway. What genuinely impresses me is how fast Anthropic ships new features 🚀

Experimenting with autoresearch — built automated YouTube Shorts generation using Remotion + Claude Code. The loop: generate → upload → analyze vs competitors → optimize for views. New channel:

Went deeper into Karpathy autoresearch. 3 parts: • Static: data, functions, optimization • Dynamic: strategies, models, approaches • program.md: full workflow description All fits in agent context → it understands. Testing: crypto algotrading + ad CTR 🚀

Everyone's excited about "auto-research" after Karpathy's post. But for anyone who ran ML experiments or competed on Kaggle — this is just a basic optimization loop. We've been doing this for years. Why the sudden hype? 🤔

T3 Code — a third-party automation platform for AI coding agents. • Codex: supported ✅ • Claude Code integration: coming soon Managing AI dev tools is becoming its own category 🤖

Customized last30days-skill for instant research across Reddit, X, YouTube, HN & Polymarket 🔍 Synthesizes social discourse into summaries — great for understanding what people actually think about any topic before making product or content decisions.

My AI assistant flagged this in the morning briefing: https://t.co/8BOeWiSkw9 Infrastructure for background agents — describe a task, forget about it. The agent ships the feature, deploys to prod, opens a PR. Deep research too. Concept looks genuinely interesting.

Found a bug in my Claude Code automation 🐛 Sessions showed 'tool loaded' instead of command names — impossible to track cron jobs. Fixed with a GitHub issue workaround ✅ Note: session IDs aren't accessible inside Claude Code. Workaround — save via hooks to temp files.

Claude Code now has an official Setup plugin 🔌 Analyzes your codebase and recommends tailored automations: hooks, skills, MCP servers, and subagents for workflow enhancement.

Applied for the Anthropic ambassador role in Belgrade. Let's see how it goes 😄

New Claude Code skill: generate Excalidraw diagrams 🎨 Now you can ask Claude to sketch architecture, flows, or any diagram — in Excalidraw's hand-drawn style. Right from the terminal. #ClaudeCode #Excalidraw #AI #DevTools

After many attempts, my automation finally works for at least one project, one team 🎉 Every morning at 8 AM it runs and finishes with suggestions. I check results with `claude -r`, review the recommendations, say "Let's do it" — and it does. Progress! 😃

Last few days: unifying workflows across products & teams 🔧 Built agent approval flows via MCP with oversight controls — automation that keeps humans in the loop where it matters. Quality > speed when agents act on your behalf 🤖