Introducing HyperFrames Open-Source Video Framework

𝐒𝐩𝐞𝐧𝐭 𝐥𝐚𝐬𝐭 𝐰𝐞𝐞𝐤𝐞𝐧𝐝 𝐛𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐚 𝐬𝐜𝐫𝐞𝐞𝐧𝐬𝐡𝐨𝐭-𝐭𝐨-𝐇𝐓𝐌𝐋 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 𝐨𝐧 𝐊𝐢𝐦𝐢 𝐊2.6. 𝐓𝐡𝐞 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧: 𝐜𝐚𝐧 𝐚 𝐬𝐢𝐧𝐠𝐥𝐞 𝐨𝐩𝐞𝐧 𝐦𝐨𝐝𝐞𝐥 𝐜𝐥𝐨𝐬𝐞 𝐭𝐡𝐞 𝐰𝐡𝐨𝐥𝐞 𝐥𝐨𝐨𝐩 - 𝐯𝐢𝐬𝐢𝐨𝐧, 𝐜𝐨𝐝𝐞, 𝐬𝐞𝐥𝐟-𝐜𝐫𝐢𝐭𝐢𝐪𝐮𝐞 -𝐰𝐢𝐭𝐡𝐨𝐮𝐭 𝐬𝐭𝐢𝐭𝐜𝐡𝐢𝐧𝐠 𝐭𝐰𝐨 𝐨𝐫 𝐭𝐡𝐫𝐞𝐞 𝐀𝐏𝐈𝐬 𝐭𝐨𝐠𝐞𝐭𝐡𝐞𝐫? Mostly yes. With one surprise. K2.6 reads the screenshot, emits a structured spec, generates self-contained HTML, renders it in headless Chrome, then critiques its own output and iterates. Recreates the source page from a single image - typography, layout, status chips, the small details. No design files. No fine-tuning. Two findings if you're building agentic flows on K2.6: 1. Thinking mode breaks the generator stage. The model burns its token budget on internal reasoning before finishing the HTML - output stops mid-tag. Instant mode fixes it. 2. The 256K context is the real unlock. Target image, generated HTML, and critique fit in a single critique call without summarisation. Most open vision models force you to compress between stages and lose fidelity. 𝐖𝐨𝐫𝐭𝐡 𝐧𝐨𝐭𝐢𝐧𝐠: K2.6 isn’t unique on every axis - GLM-5.1 leads front-end generation, Claude and GPT-5 would handle the loop on the closed side. What’s specific to K2.6 is its training for long-horizon agentic stability - the property of staying coherent across many iterations as context grows. That’s the property a critique loop ultimately depends on as it scales. Wrapped as a CLI and an MCP server (FastMCP, stdio) - Claude Desktop, Cursor, Windsurf can call it as a tool. 𝐒𝐭𝐚𝐜𝐤: Python 3.12, Playwright, Pillow, mcp[cli]. ~700 LOC. If you've made critique loops actually converge in production agents, would genuinely like to compare notes - that's the part with the most room left to improve. 𝐑𝐞𝐩𝐨: https://lnkd.in/gF3-srS5 #KimiK2 #MCP #AIAgents #OpenSource

🚨 BREAKING: HeyGen just open-sourced the video framework the entire AI agent ecosystem has been missing. It's called Hyperframes. HTML in. MP4 out. Built from day one for agents. Every other video creation framework has the same problem. They were built for humans with mouse cursors. AI agents can't drag a clip. Can't scrub a timeline. Can't click a keyframe. But they can write HTML. So Hyperframes uses HTML as the entire composition format. Data attributes define timing. Elements define layers. The browser renders it. FFmpeg encodes it. Fully deterministic same HTML produces identical MP4 output every single time. The agent skills are what make this production-ready. Hyperframes ships skills for Claude Code, Cursor, Gemini CLI, and Codex that encode framework-specific patterns how to structure compositions, write captions, sequence GSAP animations correctly. Not generic HTML docs. Not Stack Overflow answers. Actual Hyperframes patterns that work. Install automatically on project init: npx hyperframes init my-video Then your agent knows how to use it before writing a single line. Full package breakdown: → CLI: create, preview, lint, render → Core: types, parsers, linter, frame adapters → Engine: Puppeteer capture + FFmpeg encode → Producer: full pipeline with audio mixing → Studio: browser-based editor UI Built by HeyGen. 100% Open Source. Apache 2.0 License. https://t.co/moyNvAVelP

🐍 Most agencies upload unoptimized images and call it done. We run every image through a custom Python + Pillow pipeline before deployment: ✅ EXIF metadata stripped (hidden data slows loads) ✅ Resized to exact viewport dimensions ✅ Converted to WebP format ✅ Progressive loading enabled Images that enhance design without dragging down speed. Every contractor website we deploy in San Bernardino County and across SoCal ships with this treatment — no plugin required. 🔗 Full tech post in comments 👇 #CloudflareWorkers #AIWebsite #SchemaMarkup #MenifeeCA #ContractorWebsite #SoCal #WebDev #LocalSEO

Sharing a small tool that I built because I needed it and you might need it too. In the era of agentic development, everything is a Markdown file—AI plans, instructions, skills, and workflows are everywhere. Chrome still displays them as raw text, and downloading them from Slack or GSuite just to read them is a total flow-killer. So I built MarkUp. It’s a Chrome Extension that renders .md files beautifully, right in the browser. The Essentials: * Smart Rendering: Intercepts web downloads and renders local .md files instantly. * Full Features: Themes, TOC, search, and syntax highlighting built-in, and many more. * Clean: Zero data collection. No dependencies. Pure vanilla JS. The project is fully open source and easy to set up. Checkout here: https://lnkd.in/gz_ayzpM #AI #AgenticWorkflows #DeveloperTools #JavaScript

NEW: @liu8in at HeyGen just open-sourced HyperFrames and AI agents can now edit videos. HyperFrames turns browser-native code (HTML, CSS, GSAP) into rendered video frames. Agents already know the web and now they can ship videos from it. Zero API keys. Fully local. Fully open source.

Cursor 3 just dropped and it's not an update. They've rebuilt the entire thing from scratch. The old Cursor had become clunky. Full of questionable design choices. It was starting to feel like a vibe coded slot machine where you'd pull the lever and hope for decent output. I'd mostly moved away from it. But Cursor 3 is a completely different application. The fundamental shift is that it's no longer built around files. It's built around agents. The UI is stripped back. Projects and chats sit in the sidebar. You can have multiple projects open with different agents running simultaneously. There's a new design mode (Shift+Command+D) that lets you visually select elements and draw on screen, sending screenshots directly to the model. Here's what's interesting though. If you look at Cursor 3 next to OpenAI's Codex app and T3 Code (Theo's new editor), they all look remarkably similar. The entire industry is converging on the same agent-first UI pattern at exactly the same time. I've actually been preferring T3 Code lately because its free, open source, and works with my existing Claude and Codex subscriptions rather than requiring another monthly payment. But the bigger picture matters more than which tool wins. We're watching a spectrum of control emerge in real time. On one end you've got VS Code where you have full manual control. Then Copilot gives you inline suggestions. Cursor puts agents right in front of you. And at the far end, pure agent mode where you never even look at code. There's a genuine tension here that nobody's resolved yet. The further you move toward agent-first workflows, the less aware you are of what's actually happening in your codebase. You gain speed but lose visibility. Claude Code in the terminal still feels more transparent to me. You can see exactly which lines are changing and why. With these new agent-first editors, you're trusting more and verifying less. The question isn't which tool is best. It's how much control are you willing to give up, and what are you getting in return. Read more: https://lnkd.in/esc-EFsq

BREAKING: HeyGen just open-sourced the video framework the entire AI agent ecosystem has been missing. It's called Hyperframes. HTML in. MP4 out. Built from day one for agents. Every other video creation framework has the same problem. They were built for humans with mouse cursors. AI agents can't drag a clip. Can't scrub a timeline. Can't click a keyframe. But they can write HTML. So Hyperframes uses HTML as the entire composition format. Data attributes define timing. Elements define layers. The browser renders it. FFmpeg encodes it. Fully deterministic same HTML produces identical MP4 output every single time. The agent skills are what make this production-ready. Hyperframes ships skills for Claude Code, Cursor, Gemini CLI, and Codex that encode framework-specific patterns how to structure compositions, write captions, sequence GSAP animations correctly. Not generic HTML docs. Not Stack Overflow answers. Actual Hyperframes patterns that work. Install automatically on project init: npx hyperframes init my-video Then your agent knows how to use it before writing a single line. Full package breakdown: → CLI: create, preview, lint, render → Core: types, parsers, linter, frame adapters → Engine: Puppeteer capture + FFmpeg encode → Producer: full pipeline with audio mixing → Studio: browser-based editor UI Built by HeyGen. 100% Open Source. Apache 2.0 License. Repo Link In Comments

🐍 Unoptimized images are the #1 cause of slow contractor websites. We don't use compression plugins. We built a custom Python + Pillow pipeline that: → Strips all hidden EXIF metadata → Resizes to exact viewport dimensions → Converts to WebP with progressive loading Every image on every site we deploy is processed before it ever hits the server. That's how we hit 90+ PageSpeed scores in Menifee and across all of Southern California. 🔗 Full stack post in comments 👇 #WebDesign #LocalSEO #MenifeeCA #SoCal #ContractorMarketing #TechStack #HTMLWebsite #PageSpeed

🐍 Unoptimized images are the #1 cause of slow contractor websites. We don't use compression plugins. We built a custom Python + Pillow pipeline that: → Strips all hidden EXIF metadata → Resizes to exact viewport dimensions → Converts to WebP with progressive loading Every image on every site we deploy is processed before it ever hits the server. That's how we hit 90+ PageSpeed scores in Menifee and across all of Southern California. 🔗 Full stack post in comments 👇 #WebDesign #LocalSEO #MenifeeCA #SoCal #ContractorMarketing #TechStack #HTMLWebsite #PageSpeed

🚀 The Cursor 3.0 Power-User Guide While .cursorrules and indexing remain the bedrock, 3.0 introduces a paradigm shift: Multi-agent autonomy. 1. Master the "Agents Window" (Cmd+Shift+P → Agents) The 3.0 update moved the AI from a sidebar chat to a dedicated Agents Window. This allows you to run multiple agents in parallel across different repositories. ◾ The Pro Move: Use Cloud Agents for long-running tasks. You can start a massive refactor, close your laptop, and the agent continues running in an isolated Ubuntu VM. When you return, just review the diffs and click "Apply." 2. Design Mode & The Integrated Browser Cursor 3.0 isn't just for logic; it’s now visually aware. ◾ How to use it: Open the built-in browser and toggle Design Mode (⌘ + Shift + D). ◾ The Efficiency Hack: Instead of describing a bug, just highlight the UI element in the browser. Cursor can capture console errors, network traces, and DOM nodes automatically to fix CSS or functional bugs. 3. Leverage "Composer 2" (The New Engine) Cursor recently launched Composer 2, a model specifically trained on "long-horizon" coding tasks. ◾ Why it matters: Unlike standard LLMs, Composer 2 is optimized for tasks requiring hundreds of steps. Use it for complex features like "Add a complete Stripe subscription flow" rather than just single-file edits. 4. Advanced Context Control: The @ System 2.0 You mentioned Reference Open Editors, but 3.0 expanded the @ symbol significantly: ◾ @Web: Fetches the latest docs (crucial for fast-moving libraries like Next.js or LangChain). ◾ @Codebase: Scans your entire indexed project to find where a logic pattern is used. ◾ @Git: Mentions specific commits or diffs to explain why a bug was introduced. The "Cursor Philosophy" for 2026: 💡 Stop thinking about "asking questions" and start thinking about "assigning tasks." With 3.0, you aren't just getting code snippets; you’re managing a fleet of agents that can navigate your terminal, browser, and multi-repo architecture. Which of these features are you planning to integrate into your workflow first?