GitHub 2026-04-25 Weekly 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.

A Claude Code skill that automates decompilation of Android packages and extracts Retrofit/OkHttp HTTP APIs, aimed at security audits, interoperability analysis, and incident response.

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

Claude Context is an MCP-based semantic code-search plugin that vectorizes your code and stores it in a vector database, injecting only relevant snippets into Claude's context on demand; it reduces invocation cost and improves relevance for large-codebase AI coding workflows, but requires external embedding and vector DB services and lacks clear license/maintenance visibility.

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.

For teams needing auditable, protocolized prompt evolution, Evolver turns log-driven prompt tweaks into standardized, reusable GEP assets and records EvolutionEvents; it supports offline use with optional network integration.

OpenAI Agents SDK provides a lightweight, extensible multi-agent workflow framework with multi-LLM support, configurable guardrails, and built-in tracing—suited for rapid prototyping and complex dialogue orchestration.

OpenSRE offers a self-hostable framework for AI SRE agents with training and evaluation environments, enabling evidence-backed incident investigation and automated remediation in real cloud scenarios — suited for teams embedding LLM capabilities into production operations.

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.

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.

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.

A context-management server for MCP and LLM toolchains: it sandboxes outputs, indexes events with SQLite+FTS5 and retrieves via BM25, drastically cutting context usage while preserving session continuity—suitable for teams needing token-cost control and persistent conversation state.

DeepGEMM is a high-performance CUDA kernel library for NVIDIA SM90/SM100 that delivers JIT-compiled FP8/FP4/BF16 GEMMs and fused MoE/MQA primitives, targeting LLM training and inference workloads that demand extreme throughput and low latency; however, it imposes strict hardware/CUDA requirements and has unresolved licensing and community-sustainability concerns.

Voicebox is a local-first, API-first open-source voice cloning and synthesis studio that provides a complete toolchain for developers and creators needing privacy, local deployment, multi-language support, and customizable audio effects.

Dive into LLMs delivers systematic, hands-on tutorials for students and researchers covering fine-tuning, prompting, RLHF, model safety and multimodality, with code-focused, course-ready materials.