Today's trending projects tell a coherent story: the AI agent ecosystem is maturing past single-model experiments and into durable infrastructure. Developers are building platforms that handle multiple models, multiple channels, and multiple task types — and the common thread is interoperability. The question is no longer whether an AI can do a given task; it's how to wire it cleanly into the rest of the stack.
Hermes Agent: A Self-Improving, Model-Agnostic Agent Platform
NousResearch's hermes-agent sits at 172,000 stars and added over 1,400 today. The pitch is direct: one agent platform that works across 200+ models (OpenAI, Anthropic, Hugging Face, and Nous's own portal) and every communication channel you already use — Telegram, Discord, Slack, WhatsApp, Signal, and CLI through a unified gateway. The platform creates and refines its own skills autonomously, runs scheduled tasks via a built-in cron system, and adheres to the agentskills.io open standard for procedural memory, which means skills built here are portable across other compliant platforms. It runs on a laptop, in Docker, over SSH, or on Modal serverless functions, keeping the hosting barrier low.
github.com/NousResearch/hermes-agent
Anthropic's Public Skills Repository Crosses 143K Stars
The anthropics/skills repository — Anthropic's reference implementation for the Agent Skills standard — keeps gaining momentum, adding 718 stars today on its way past 143,000 total. Skills in this system are YAML + Markdown files that teach Claude how to complete specific tasks repeatably: document manipulation across PDF, Word, Excel, and PowerPoint formats; web app testing; MCP server generation; creative workflows; and business processes. The repo also contains the Agent Skills specification itself, which is the same spec Hermes Agent and other multi-platform tools reference. For anyone building with Claude, this is both a working pattern library and the closest thing to official documentation on how Anthropic thinks agent memory should be structured.
MarkItDown: Microsoft's Document Pipeline for LLMs
microsoft/markitdown added another 1,400 stars today, reaching 128,000 total, and the use case is precise: getting structured document content into language models without losing the outline. The tool converts PDFs, Word docs, PowerPoint decks, Excel spreadsheets, images (with OCR), audio files (with transcription), HTML, CSV, XML, and more into clean Markdown — preserving headings, tables, and lists while stripping visual formatting that models can't use anyway. It ships as a CLI, a Python library, and a Docker image, and supports plugin extensions for less common formats. For any pipeline that needs to feed office documents into an LLM reliably, it solves a genuinely tedious problem.
github.com/microsoft/markitdown
ECC: One Configuration Layer Across All AI Harnesses
affaan-m/ECC picked up 1,385 stars today on its way to 198,000+ total. The core problem it targets is fragmentation: separate configuration setups for Claude Code, Cursor, OpenCode, Codex, and other harnesses mean rebuilding the same tooling primitives on every platform. ECC addresses this with 63 specialized agents for delegated tasks, 249 reusable skills across frameworks and languages, configurable hooks across 15+ event types, and AgentShield security scanning covering over 1,200 test cases. Language-specific rules span TypeScript, Python, Go, Java, Kotlin, and Swift. Selective installation keeps context windows from bloating — you pull only the components relevant to your current stack.
Also Trending
harry0703/MoneyPrinterTurbo added nearly 4,700 stars in a single day — one of the biggest jumps on GitHub this week — continuing its run as the most-watched AI video generation tool on the platform. It produces short-form social video end-to-end from a text prompt, handling script, visuals, voice-over, and editing in one step.