GitHub 2026-05-04 Monthly Trending Projects

Condenses Andrej Karpathy's observations about LLM coding pitfalls into four installable CLAUDE.md principles—explicit assumptions, simplicity-first, surgical changes, and goal-driven execution—helping teams limit unnecessary edits and make LLM-driven changes verifiable.

Hermes Agent is a server-deployable autonomous AI agent offering a terminal-first interface, multi-platform messaging gateway, persistent memory and skills hub, sandboxed terminals, and scheduling—suitable for long-running, auditable, and extensible automation in engineering and operations.

Agent Skills packages senior-engineer workflows into reusable agent-invokable skills to improve consistency and engineering quality across development stages.

Multica delivers a self-hosted or cloud-managed platform to turn coding agents into assignable, monitored teammates with reusable skills—aimed at engineering teams seeking controlled automation.

An extensible collection of agent skills for product and engineering teams that automates PRDs, planning, issue breakdowns and TDD flows, enabling modular integration of agent capabilities into repositories.

Free drop-in Claude proxy for developers and researchers, supporting NVIDIA NIM, OpenRouter and local backends; enables seamless replacement and quota-saving optimizations.

Kronos is the first open-source foundation model for financial K-line sequences, combining hierarchical discretization and an autoregressive Transformer to provide reproducible baseline tools for quant research and probabilistic asset forecasting.

DeepTutor centers on an agent-native architecture that combines TutorBots, persistent memory, and RAG knowledge bases to deliver an extensible personalized tutoring and collaborative workspace—well suited for autonomous learning, teaching automation, and knowledge-driven conversational education products.

Archon is a knowledge and task management platform for AI coding assistants, offering an MCP interface, document retrieval and task collaboration to unify agent context and improve model-assisted development outcomes.

MarkItDown converts varied documents into structured Markdown for LLM pipelines, enabling automated text workflows while requiring care with optional dependencies and plugin compatibility.

GenericAgent is a lightweight, self-evolving local autonomous agent framework that converts complex tasks into reusable skills via ~3K core lines and nine atomic tools; ideal for advanced users needing system-level control and deep customization.

AI Edge Gallery is an on-device offline generative-AI sandbox for mobile, supporting local Gemma 4 inference, model management, and multimodal interaction—suitable for privacy-sensitive development and performance evaluation.

For developers and teams using Claude Code: automatically capture, semantically compress, and persist session outputs, then inject relevant context into future sessions to improve continuity and retrieval efficiency.

Z4nzu/hackingtool is a curated, all‑in‑one toolkit (185+ tools) for pentesters and security researchers, offering one‑liner install, Docker support and categorized management for rapid lab setup — but pay close attention to licensing and supply‑chain security.

awesome-codex-skills is a curated set of modular Codex 'skills' for the Composio/Codex ecosystem, providing SKILL packages and an installer to add actionable automations (email, issues, Slack, 1000+ apps) to the CLI/API, enabling teams to operationalize AI-driven workflows; validate license and maintenance before adoption.

Pixelle‑Video targets content creators and small teams with a modular AI short‑video pipeline: input a topic to automatically produce script, visuals, voice and final composition — suited for rapid social‑media videos and bulk content production.

RTK intelligently filters and aggregates shell command output to remove redundant content sent to LLMs, substantially lowering token costs — ideal for frequent AI-driven development and CI workflows, but license and maintenance uncertainties pose adoption and compliance risks.

LiteRT-LM is Google's edge-focused large-model inference framework emphasizing multimodality and hardware acceleration, suited for deploying high-performance offline or near-edge inference on mobile, desktop, and embedded devices.