LangChain is a comprehensive framework for building LLM-powered applications. It provides core abstractions for language models, prompts, chains, agents, memory, and tools, enabling developers to create complex AI systems like chatbots, question-answering agents, and retrieval-augmented generation pipelines. The architecture is modular with clear separation between core langchain_core foundations, legacy langchain_classic components (many deprecated), and partner integrations for specific LLM providers and vector stores. Key entry points include agent creation functions, chain constructors, and memory/retriever interfaces. The framework emphasizes composability through the Runnable interface (LCEL) and supports both synchronous and asynchronous operations. While extensive, the codebase shows strong maintenance practices with active dependency management and version consistency checks across integrations.
Reasonix is a local AI coding agent engine written in Go, providing a terminal UI, desktop application (Electron + React), and HTTP server for interacting with large language models (OpenAI, Anthropic, DeepSeek, etc.) to automate software engineering tasks. The kernel manages sessions, context windows, planning, sub-agent delegation, tool execution (file operations, shell, browser, MCP servers), and a sophisticated safety sandbox. It persists conversation history with checkpoints for undo/rewind, tracks token usage, and maintains an evidence ledger for auditing tool effects. The architecture is a monolithic Go binary with multiple frontends: a CLI/TUI, a desktop host that communicates via SSE/JSON, and a REST API server. A plugin system using the Model Context Protocol (MCP) allows extension. Separate Cloudflare Workers handle account management, forum, and crash reporting. The codebase is well-structured with a clean layered architecture (base utilities, domain contracts, state persistence, runtime agent, assembly wiring, frontends), extensive test coverage, and internal linters enforcing quality. It is actively maintained with frequent commits and professional CI/CD pipelines.
Nanobot is a self-hosted, open-source personal AI assistant framework that operates as a full-stack application. The Python backend provides an agent loop with tool execution, session management, multi-provider LLM orchestration (OpenAI, Anthropic, etc.), and a plugin system for 18+ messaging channels (Telegram, Discord, Slack, WeChat, Feishu, etc.). It exposes a WebSocket-based gateway, an OpenAI-compatible API, and a CLI. Two frontends exist: a rich WebUI built with React/TypeScript/Vite/Tailwind and a terminal TUI built with Bun/OpenTUI. The architecture is event-driven via a message bus, with decoupled channel adapters, provider abstraction, and a config system supporting hot-reload. It includes security features like workspace sandboxing, credential redaction, and SSRF protection. The codebase is extensively tested (hundreds of unit/integration tests) and demonstrates strong engineering practices with clean abstractions, though its large dependency surface and multiple frontends introduce some maintenance surface area.
Neuron AI is a PHP framework for building AI-powered applications with support for multiple large language model providers (OpenAI, Anthropic, Gemini, Mistral, Ollama, AWS Bedrock, etc.), tool integration, structured output, retrieval-augmented generation (RAG), and complex workflow orchestration. It provides a composable agent class that can operate in chat, stream, and structured output modes, a workflow engine with interrupt/resume and persistence, a built-in evaluation framework with LLM-based judges and assertions, and a Model Context Protocol (MCP) client for connecting to external tool servers. The architecture is modular: agents are built on top of workflows, tools are defined via PHP classes with JSON schema generation, and providers are abstracted behind a common interface with traits for shared logic.
OpenRath is a framework for building LLM-powered AI agent applications with a torch-like API design. It centers on the Session abstraction (analogous to a Tensor) that flows through Sandboxes (execution environments), Memories (persistent state), and Tools (callable operations), orchestrated by Agents and Workflows. The framework provides session lineage tracking, sandboxed tool execution, memory persistence, and LLM provider abstraction, enabling the creation of multi-agent workflows with durable state and traceable provenance. The architecture is modular and extensible, featuring pluggable backends for sandboxes (local, OpenSandbox) and memory systems (local, OpenViking), a registry pattern for dynamic backend loading, and lazy evaluation of sessions for performance. Key innovations include write-ahead logging for crash-resistant session persistence, resource-keyed parallelism for safe concurrent tool execution, and unified garbage collection across session, sandbox, and memory domains. The framework is designed for production use, with extensive testing, clear documentation, and active maintenance.
OpenManus is an open-source framework for building general AI agents, inspired by the Manus agent. It provides a modular architecture with a hierarchy of agents (BaseAgent, ReActAgent, ToolCallAgent, Manus, SWEAgent, BrowserAgent, etc.), a rich set of tools (bash, file editing, web search, browser automation, code execution, chart visualization, desktop automation), and multiple execution modes: single-agent CLI, multi-agent planning flow, MCP client/server, and sandboxed execution via Docker or Daytona. The system is LLM-agnostic, supporting OpenAI, Anthropic, Azure, AWS Bedrock, Ollama, and others through a configurable LLM abstraction. It also implements the A2A (Agent-to-Agent) protocol for inter-agent communication. While primarily an application framework, it includes data analysis and chart generation capabilities, giving it a hybrid character. The codebase is well-structured with clean abstractions, but contains some hardcoded values and Chinese comments that may hinder maintainability.
Swarm is a production-grade Swift-native agent runtime framework for building multi-agent conversational systems with type-safe tools, on-device Apple Foundation Models, composable workflows, memory, guardrails, and observability. The architecture is modular: core agent execution (Agent, Workflow, Job) with pluggable inference providers (Foundation Models, OpenAI-compatible, mock), memory backends (conversation, sliding window, summary, vector, persistent via SwiftData or Wax SQLite), and tool integration (built-in tools, MCP protocol, web search). Key entry points include Agent.run()/stream(), Workflow.execute(), and the SwarmMCPServerService for exposing tools via MCP. Data flows involve agent turns with tool calls, memory retrieval/consolidation, and durable workflow checkpointing via the HiveCore deterministic runtime. External dependencies include Apple FoundationModels (macOS 26+), swift-log, swift-syntax (for macros), and optional Wax (SQLite vector store) and OpenTelemetry SDK. The framework emphasizes compile-time safety via @Tool and @Agent macros, and provides resilience patterns (retry, fallback, circuit breaker, rate limiting) and observability (tracing, metrics, OpenTelemetry).
Dexter is a CLI-based AI agent framework specifically designed for deep financial research. It implements a sophisticated agent loop that combines task planning, autonomous tool execution, and self-validation to answer complex financial queries. The system features a modular architecture with specialized systems for memory management, tool execution, permissions, and skills. Key capabilities include access to real-time financial data through multiple APIs, web search functionality, file system operations, browser automation, and specialized workflows like DCF valuation. The agent maintains persistent memory across sessions using vector embeddings and keyword search with temporal decay weighting. Dexter can be deployed via CLI or as a WhatsApp bot, and includes an evaluation framework for testing agent performance against financial questions. The system emphasizes security through a robust permission system that controls access to potentially dangerous operations like shell commands.
PrivateGPT is a production-grade, private AI assistant platform that provides RAG (Retrieval-Augmented Generation) over ingested documents, tool/function calling, sandboxed code execution, web search, MCP (Model Context Protocol) integration, skill management, and multimodal support. It is built as a FastAPI server with background workers (Celery and ARQ) for asynchronous ingestion and task processing, and exposes a Claude API-compatible streaming interface. The architecture is modular: a core chat engine with an interceptor pipeline for request/response processing, a pluggable component system with registries for LLMs, embeddings, readers, vector stores, and sandboxes, and a two-tier scheduling system (local and distributed) for tool and chat execution. The codebase demonstrates strong engineering practices with dependency injection, registry patterns, and composable interceptors, but its dual scheduling backends, extensive monkey-patching of third-party libraries, and heavy reliance on llama-index introduce maintenance complexity and technical drift. Reusable highlights include the context stack builder for dynamic system prompt assembly, the interceptor chain pattern, the sandbox abstraction for secure code execution, and the citation parser with incremental streaming support. The project is actively maintained by Zylon and powers their enterprise offering, with 57k+ stars and frequent releases.
CowAgent is a production-grade multi-agent AI assistant platform that routes messages from 12+ IM channels (WeChat, Feishu, Telegram, DingTalk, Slack, Discord, QQ, WeCom, web console, desktop Electron app) through a modular agent core. The core implements planning, tool use (bash, file editing, browser, web search, MCP, sub-agent delegation, scheduler, memory, vision), three-tier memory with vector embeddings and markdown-aware chunking, knowledge base management, skill system with SKILL.md manifests and a skill hub, self-evolution (idle-conversation review loop), and multi-agent team orchestration with peer transport. The architecture is a modular monolith: channel adapters normalize messages into a unified ChatMessage, the Bridge dispatches to per-agent/session model resolution, the Agent protocol (Task/AgentResult/AgentAction) drives the execution loop with streaming, cancellation, and steering, and all state persists to SQLite with incremental migration. Key entry points are app.py (startup orchestration), cli/cli.py (daemon lifecycle), channel/web/web_channel.py (HTTP API and SSE streaming), and the desktop Electron client (React/TypeScript with Zustand stores). External dependencies include LLM APIs (OpenAI, Claude, Gemini, DeepSeek, etc.), IM platform webhooks/SDKs, optional voice TTS/ASR providers, and MCP servers. The codebase demonstrates strong engineering practices with extensive regression tests (200+ test files), atomic file writes, thread-safe context propagation via ContextVar, and security guards for credential paths and SSRF.
AG2 is a production-grade Python framework for building multi-agent AI systems. It provides a unified agent abstraction with support for multiple LLM providers (OpenAI, Anthropic, Gemini, Mistral, Bedrock, Ollama, xAI, Z.AI, DashScope), protocol integrations (A2A, A2UI, MCP, ACP), and sandboxed code execution (Docker, Daytona, local). The architecture is event-driven with a modular middleware pipeline, dependency injection, and a streaming event bus. It includes a full evaluation suite with trace reconstruction, scorers, pairwise comparison, and result persistence. The network layer enables distributed agent coordination via a hub-and-channel model with WAL persistence, identity management, and transport-agnostic envelopes. The framework also supports real-time voice interaction, skills-based toolkits, knowledge stores, and human-in-the-loop workflows. While the surface area is large, the codebase demonstrates high craftsmanship with clean abstractions, protocol-based interfaces, and extensive test coverage across all modules.
napari-chatgpt (Omega) is an LLM-powered autonomous agent designed as a napari plugin for interactive image analysis. It enables users to interact with the napari viewer via natural language through a chat interface, either embedded in napari or via a web-based UI. The system processes user commands by leveraging LLMs (via LiteMind) to select and execute appropriate tools for tasks such as image loading, segmentation, denoising, widget creation, viewer control, web search, and code execution. The architecture centers around a plugin-contributed dock widget that launches a chat interface, which communicates with a FastAPI/WebSocket backend. The backend hosts the Omega agent, which uses a modular tool system where each tool (e.g., for file operations, segmentation, or web requests) inherits from a base class and is responsible for executing specific actions. The agent can iteratively refine its actions through dialogue, enabling complex workflows like 'segment nuclei and show measurement statistics'. Key external dependencies include napari (>=0.5), LiteMind, LLM provider libraries (OpenAI, Anthropic, Google), FastAPI, Qt bindings, and scientific libraries like scikit-image, Cellpose, and StarDist.
smolagents is a HuggingFace library for building LLM-powered agents, with a strong emphasis on 'code agents' that generate and execute Python code as actions. The core architecture consists of a ReAct-style loop (MultiStepAgent), with two main implementations: CodeAgent (writes Python code) and ToolCallingAgent (writes JSON tool calls). Agents are model-agnostic, supporting dozens of backends (OpenAI, LiteLLM, Hugging Face Inference API, local Transformers, vLLM, MLX, Bedrock, etc.) via a clean Model abstraction. Tools are defined as Python classes with typed inputs/outputs and automatic JSON schema generation, or via a simple @tool decorator, and can be shared on the HuggingFace Hub. The library includes a built-in AST-based Python executor (LocalPythonExecutor) for safe agent code execution, though it is explicitly not a security boundary; production use is directed toward sandboxed remote executors (E2B, Docker, Modal, Blaxel). Memory is managed through structured step objects (ActionStep, PlanningStep, TaskStep) with callbacks for customization. Additional features include MCP client integration, Gradio-based interactive UI, CLI tools, telemetry via OpenTelemetry, and serialization with SafeSerializer.
GEM is a Gym-like framework for training and evaluating agentic LLMs via reinforcement learning. It provides a diverse set of environments (games, math, code, QA, etc.) that can be used to train LLMs to interact with tools, solve problems, and make sequential decisions. The framework features a modular design with composable wrappers for observation modification, tool integration, and episode tracking, along with vectorized environments for parallel execution. It supports integration with multiple RL training frameworks (Oat, Verl, Tinker, OpenRLHF, RL2) through example scripts, making it a versatile tool for LLM agent research.
Debug-gym is a framework for developing and evaluating AI agents that perform interactive debugging of software. It provides a Gymnasium-inspired environment where agents can use tools like a debugger (pdb), file viewer, editor, and shell to inspect and modify code in order to fix bugs. The framework supports multiple execution backends (local, Docker, Kubernetes) to ensure isolation and reproducibility, and integrates with various LLM providers through a unified interface. The system is designed around a loop of observation, action, and feedback: the agent observes the current state (e.g., code, test output), selects a tool to apply (e.g., set a breakpoint, edit a file), and receives the outcome from the environment. This enables the creation of agents that can learn to debug or be evaluated on debugging benchmarks such as SWE-Bench and Mini-Nightmare. The framework also includes analysis tools for visualizing agent trajectories and logging experiment results, supporting both development and research use cases. Key architectural features include a plugin-based tool system, abstract terminal and LLM interfaces, and a secure workspace for code manipulation. These components are designed to be modular and extensible, allowing researchers and developers to plug in new tools, environments, or language models with minimal changes to the core loop.
AgentFlow is a framework for building and training agentic LLM systems using reinforcement learning, specifically Flow-GRPO (Flow-based Group Refined Policy Optimization). It provides a modular architecture with four core roles—Planner, Executor, Verifier, and Generator—that coordinate via evolving memory and tool calls over multiple turns. The system supports both inference (via a Solver) and distributed RL training (via a Trainer/Client/Server architecture built on verl, vLLM, and Ray). It includes a multi-provider LLM engine abstraction (OpenAI, Anthropic, Gemini, DeepSeek, etc.), a tool system with metadata-driven discovery, and instrumentation for tracing and reward collection. The repository is heavily research-oriented, with benchmark datasets (AIME24, GPQA, GAIA, HotPotQA, etc.) and training pipelines for optimizing agent planners. While primarily an application framework, it has strong data/research components, making it a hybrid project with a focus on reproducible agentic reasoning experiments.
Pixel-Reasoner implements a two-stage training framework for Vision-Language Models (VLMs) to enable pixel-space reasoning. The system first performs instruction tuning on synthesized reasoning traces to familiarize models with visual operations (zoom-in, select-frame), then applies curiosity-driven reinforcement learning to balance exploration between pixel-space and textual reasoning. Targeting visual reasoning benchmarks (V* bench, TallyQA-Complex, InfographicsVQA, MVBench), it provides pre-trained models and evaluation scripts. The architecture combines adapted OpenRLHF components with custom VLM modules, Ray-based distributed training, and specialized data processors for multimodal input. Key entry points include training scripts (single/multi-node) and evaluation scripts, with data flowing from Hugging Face datasets through tokenization/image processing to PPO optimization loops.
This repository is a large-scale educational resource and community showcase for building LLM-powered agents using the custom 'HelloAgents' framework. It is organized as a multi-chapter tutorial (16 chapters) covering agent fundamentals, tool use, memory systems, multi-agent protocols (MCP/A2A/ANP), context engineering, reinforcement learning (SFT→GRPO), and evaluation. Each chapter contains runnable code examples, and the repository also hosts 40+ community-contributed 'Co-creation-projects' that demonstrate real-world agent applications across domains including stock analysis, health record management, academic research, code assistance, game design, and data analysis. The architecture is not a single deployable system but a polyglot collection of independent projects, most following a FastAPI backend + Vue/React frontend pattern, with a shared dependency on the HelloAgents framework core. The primary data flow in each project is: user input → LLM reasoning loop (ReAct/Plan-and-Solve/Reflection) → tool execution (search, database, code execution, APIs) → memory/context update → response generation. External dependencies include OpenAI-compatible LLM APIs, search engines (Tavily, Brave, DuckDuckGo), vector stores (Qdrant, Milvus, Neo4j), and various domain-specific APIs (financial data, academic databases, weather). The codebase serves as both a learning path for agent development and a practical reference for common agent patterns, though quality varies significantly between the polished tutorial chapters and the prototype-level community projects.