GitHub 2026-05-12 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.

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.

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.

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.

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.

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.

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.

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.

An OSINT-oriented username enumeration tool that automatically searches and aggregates account profiles across thousands of sites, provides multi-format reporting and proxy/Tor support, and is suited for preliminary identity linking and lead collection.

TradingAgents constructs a modular, configurable research pipeline of role-based LLM agents, supporting multi-provider models and simulated execution—suited for academic and R&D strategy validation but not investment advice.

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.

Open Generative AI is an open, uncensored multi‑modal generative studio for creators, offering hosted and desktop local inference with 200+ models; however, license status is unclear and maintenance activity appears limited, so commercial adoption requires caution.

PPT Master is a local-first open-source workflow that uses AI IDEs and large models to convert PDFs, DOCX, URLs or Markdown into natively editable PPTX files—suited for professionals who prioritize editability, data privacy, and transparent costs.

DFlash applies block diffusion to speculative decoding, using parallel draft models to boost inference throughput and response speed—suitable for high-concurrency serving and research validation; however, the repository lacks clear license, contributors, and releases, so production use requires careful evaluation.