Дайджест awesome-списков korchasa (11.06.2026)

Новый срез изменений в awesome-списках korchasa получился крупным: 183 новых пункта по трём направлениям. Основной объём пришёлся на AIOps, где добавлено 167 проектов; в AI Agents появилось 13 позиций, в MCP — 3. Ниже — не просто перечень ссылок, а контекст по каждому проекту из опубликованной выборки: что именно добавлено, насколько заметен репозиторий по звёздам и какую практическую нишу он закрывает.

AI Agents

В разделе AI Agents виден широкий спектр инфраструктуры для агентных систем: оркестрация, долговременная память, изоляция выполнения, фреймворки разработки и инструменты доступа к внешнему контенту.

• ruvnet/ruflo — 58 416 звёзд. Ruflo is a leading agent orchestration platform for Claude, enabling the deployment of intelligent multi-agent swarms, coordination of autonomous workflows, and building conversational AI systems.
• siyuan-note/siyuan — 44 321 звезда. SiYuan is a privacy-first, self-hosted, open-source personal knowledge management system with block-level referencing and Markdown WYSIWYG editing capabilities.
• Panniantong/Agent-Reach — 23 302 звезды. Agent-Reach allows AI agents to access web content from major social media and video platforms through a CLI, without API fees, enhancing their perceptual capabilities.
• NVIDIA/NemoClaw — 21 047 звёзд. NVIDIA NemoClaw is an open-source reference stack enabling the secure and sandboxed operation of AI agents like OpenClaw and Hermes within NVIDIA OpenShell environments, offering managed inference …
• chopratejas/headroom — 17 189 звёзд. Headroom is a context compression layer for AI agents, reducing token usage by 60-95% with maintained accuracy across various workloads, offered as a library, proxy, or MCP server.
• opensandbox-group/OpenSandbox — 11 389 звёзд. OpenSandbox is a general-purpose, secure, and extensible sandbox runtime for AI applications, offering multi-language SDKs and robust isolation for agent execution.
• aden-hive/hive — 10 505 звёзд. Hive is an outcome-driven agent development framework designed for evolving agents with AI evaluation, human-in-the-loop, and self-improvement capabilities.
• j3ssie/osmedeus — 6 408 звёзд. Osmedeus is a powerful and modern orchestration engine designed for reconnaissance and penetration testing workflows, written in Go.
• strands-agents/harness-sdk — 6 056 звёзд. Strands Agents is a versatile SDK for building and operating AI agents with model-driven approaches, supporting various LLMs, multi-agent systems, and production deployments.
• TencentCloud/TencentDB-Agent-Memory — 5 104 звезды. TencentDB Agent Memory provides a local, 4-tier progressive pipeline for AI agent long-term memory, enhancing performance and reducing token usage without external API dependencies.
• dromara/liteflow — 3 757 звёзд. LiteFlow is a lightweight, modern rule engine framework that orchestrates complex business components and integrates AI Agents as first-class citizens.
• agentscope-ai/agentscope-java — 3 611 звёзд. AgentScope Java is an agent-oriented programming framework for building production-ready LLM-powered applications, offering features like ReAct reasoning, tool calling, memory management, and multi…
• kepano/obsidian-skills — 0 звёзд. Obsidian Skills provides a collection of agent skills enabling AI agents to interact with Obsidian vaults, facilitating content creation and editing of Markdown, Bases, and JSON Canvas files, as we…

MCP

В MCP-добавлениях хорошо заметна связка между агентами, памятью и корпоративными источниками данных: сжатие контекста, локальная долговременная память и интеграция Microsoft Teams через сервер протокола.

• chopratejas/headroom — 17 189 звёзд. Headroom is a context compression layer for AI agents, reducing LLM token usage by 60-95% while preserving accuracy, offered as a library, proxy, or MCP server.
• EverMind-AI/EverOS — 7 096 звёзд. EverOS is an open-source Python framework for self-evolving, portable, long-term memory for AI agents, supporting multimodal ingestion and local-first persistence with Markdown as the source of truth.
• InditexTech/mcp-teams-server — 354 звезды. MCP Teams Server is an MCP protocol server implementation enabling Microsoft Teams integration with capabilities to read, create, reply to messages, and mention members within Teams channels.

AIOps

AIOps обновился особенно сильно. В выборке есть инструменты для памяти агентов, оценки безопасности, LLMOps-наблюдаемости, RAG, DevOps-бенчмарков, инференса, vector search и аудита локальных LLM-прокси.

• ClaudioDrews/memory-os — 1 016 звёзд. Memory OS is a 7-layer memory operating system for Hermes Agent, providing persistent memory, structured facts, context injection, and a self-curating wiki, running locally with any LLM provider.
• wuyoscar/Internal-Safety-Collapse — 777 звёзд. ISC-Bench is a research project and benchmark for evaluating the safety collapse of frontier LLMs and AI agents, demonstrating how they can be made to produce harmful outputs under specific task co…
• stoyan-stoyanov/llmflows — 706 звёзд. LLMFlows is a Python framework for building, operating, and observing LLM-powered applications, emphasizing explicit workflows, transparency, and easy debugging.
• bosun-ai/swiftide — 704 звезды. Swiftide is a Rust library for building fast, streaming LLM applications with support for RAG pipelines, agentic workflows, and integrations with popular LLMs and storage.
• youssofal/MTPLX — 695 звёзд. MTPLX is an Apple Silicon-native inference engine that leverages Multi-Token Prediction (MTP) for speculative decoding, significantly boosting LLM inference speeds with mathematical correctness.
• ServerlessLLM/ServerlessLLM — 685 звёзд. ServerlessLLM is a system for deploying multiple AI models on shared GPUs, offering ultra-fast checkpoint loading, GPU multiplexing, and unified inference with LoRA fine-tuning.
• kossisoroyce/timber — 683 звезды. Timber is an AOT compiler that converts classical ML models (XGBoost, LightGBM, scikit-learn, CatBoost, ONNX) into native C99 code for ultra-fast, portable, and low-footprint inference serving.
• elastic/mcp-server-elasticsearch — 669 звёзд. The Elasticsearch Model Context Protocol (MCP) Server enables AI agents to interact with Elasticsearch data using natural language, query indices, and retrieve information.
• Bessouat40/RAGLight — 664 звезды. RAGLight is a modular Python framework for building Retrieval-Augmented Generation (RAG) applications, supporting various LLMs, embeddings, vector stores, and agentic workflows.
• AltimateAI/altimate-code — 661 звезда. Altimate-code is an open-source, AI-powered agentic data engineering harness for dbt, SQL, and cloud warehouses, providing deterministic SQL analysis, lineage, and FinOps.
• codefuse-ai/codefuse-devops-eval — 655 звёзд. DevOps-Eval is an industrial-first evaluation benchmark specifically designed for large language models (LLMs) in the DevOps and AIOps domains.
• toby-bridges/api-relay-audit — 655 звёзд. API Relay Audit is a security tool for local AI API relays and LLM proxies, detecting prompt injection, model substitution, tool-call rewriting, Web3 risks, and other relay tampering.
• jina-ai/vectordb — 649 звёзд. Vectordb by Jina AI is a Pythonic vector database for efficient storage, retrieval, and serving of vectorized data, with CRUD operations and scalability options including sharding and replication.
• santifer/cv-santiago — 636 звёзд. An interactive portfolio featuring an AI chatbot with agentic RAG, full LLMOps observability, automated evaluations, and prompt injection defense.
• BUAADreamer/EasyRAG — 632 звезды. EasyRAG is an efficient Retrieval-Augmented Generation (RAG) framework specifically designed for automated network operations, aiming for accurate, simple, and efficient inference.

Общее направление изменений достаточно ясное: агентные проекты всё чаще требуют не только модели и промпта, но и полноценной операционной среды — памяти, песочниц, наблюдаемости, протокольных интеграций, контроля безопасности и оптимизации стоимости инференса.