If there's a single thread running through today's GitHub trends, it's infrastructure for the agent era. Developers aren't just building with AI anymore — they're building the scaffolding that makes AI agents more capable, portable, and useful across different environments. From workflow-sharing desktops to multi-platform internet access layers to full engineering harnesses, this week's trending projects read less like demos and more like foundational tooling.
different-ai/openwork: An Open-Source Alternative to Claude Cowork
OpenWork is a cross-platform desktop app (macOS, Windows, Linux) that lets teams create AI workflows once and share them across agents — Claude Code, Cursor, Codex, and anything else that speaks MCP. It mirrors the session-sharing model of Claude Cowork but strips out the vendor dependency. The OpenWork Den control plane handles user provisioning, access control, and a skill/plugin marketplace for organizations that want central governance without locking into a single provider. With 19.1k stars and 915 gained today, the appetite for vendor-neutral workflow infrastructure is clearly real.
github.com/different-ai/openwork
affaan-m/ECC: A Full Engineering Harness for AI Coding Agents
ECC is an MIT-licensed framework that installs on top of Claude Code, Cursor, Codex, or OpenCode and gives the agent a coordinated engineering process instead of a blank slate. The harness ships with 67 specialized sub-agents (planning, security review, build repair, language-specific coding) and 281 reusable skills that load on demand rather than filling the context window on every session. Hooks handle deterministic quality checks automatically, memory persists across sessions, and an AgentShield component scans the harness itself for security risks. The core thesis: make engineering standards structural rather than conversational, so "use TDD" doesn't get lost in chat history.
mvanhorn/last30days-skill: Real-Time Social Research Across 13 Platforms
This agent skill researches any topic, person, or trend by simultaneously querying Reddit, X, YouTube, TikTok, Hacker News, Polymarket, GitHub, LinkedIn, arXiv, and more — then ranks results by actual human engagement (upvotes, views, prediction market volume) rather than SEO. Cross-platform clustering merges stories mentioned in multiple places into unified findings, and Claude synthesizes everything into a cited brief. It installs in a single command across 50+ agent platforms with zero config for Reddit, HN, Polymarket, and GitHub. The tagline is blunt: "Google aggregates editors. /last30days searches people."
github.com/mvanhorn/last30days-skill
huggingface/speech-to-speech: Modular Open-Source Voice Agent Stack
Hugging Face's speech-to-speech toolkit chains VAD, STT, LLM, and TTS into a swappable pipeline for building local voice agents. It exposes a WebSocket endpoint compatible with the OpenAI Realtime API, meaning existing integrations can point at a fully local, open-source stack without code changes. The project already powers thousands of Reachy Mini robots in production and supports Parakeet TDT for ASR, Qwen3-TTS for synthesis, and any OpenAI-compatible LLM backend. Privacy-focused teams finally have a coherent full-stack voice option that doesn't require a single API call to leave the building.
github.com/huggingface/speech-to-speech
microsoft/TRELLIS.2: Image to High-Fidelity 3D in 60 Seconds
Microsoft's 4B-parameter TRELLIS.2 converts a single image into a fully textured 3D asset — base color, roughness, metallic, and opacity included — in roughly 60 seconds on a high-end GPU. The model uses a "field-free" sparse voxel structure called O-Voxel that handles open surfaces, non-manifold geometry, and internal structures that break traditional mesh pipelines. At resolutions up to 1536³, the output is publication-ready for game assets, product visualization, or research. Pretrained weights are on Hugging Face and the repo ships training infrastructure for fine-tuning, making it more than just a demo.