An Open Stack for AI Agents
AgentScope spans the full agent lifecycle — from the core framework, to runtime hosting, to memory and evaluation, to fine-tuning, to ready-to-use applications. Following are some highlighted open-source projects that are built to work together.
Multi-language frameworks and production runtimes — the shared foundation every AgentScope project is built on.
AgentScope
Core FrameworkThe core open-source framework for building transparent, manageable AI agents — with built-in ReAct, tools, memory, planning, realtime voice, evaluation, and production-ready deployment from local to Kubernetes.
AgentScope-Java
Core FrameworkAn agent-oriented programming framework for the JVM — bringing reasoning, tool integration, memory management, and multi-agent collaboration to Java enterprise applications.
HiClaw
Multi-Agent OSAn open-source collaborative multi-agent OS for transparent, human-in-the-loop task coordination via Matrix rooms — keeping humans and agents in the same conversation.
Memory that persists, skills that compose, evaluators that judge, and training that improves — the layer that makes agents continuously smarter.
ReMe
MemoryA memory management toolkit that gives agents persistent, retrievable memory — file-based and vector-based — so they remember preferences, learn from past interactions, and stay context-aware across sessions and long conversations.
OpenJudge
EvaluationA unified framework for holistic evaluation and quality rewards, with 50+ production-grade judges that systematically measure agent lifecycle, tool use, code, math and multimodal output — and turn judgments into training rewards.
Trinity-RFT
Evolving and TrainingA general-purpose, flexible framework for reinforcement fine-tuning of LLMs. Cleanly decouples the loop into Explorer, Trainer, and Buffer — making complex RFT pipelines easy to coordinate and customize.
TuFT
Training and InfraMulti-tenant fine-tuning for local LLMs with a Tinker-compatible API — letting many users share one infrastructure to fine-tune models efficiently through a single unified interface.
Agents doing real things in real environments — a personal assistant you run yourself, a prediction arena wired to live games, a social network where agents make friends.
QwenPaw
AgentYour personal AI assistant — easy to install on your machine or the cloud, with multi-agent collaboration, memory that evolves, and integrations across DingTalk, Feishu, WeChat, Discord, Telegram and more.
DojoZero
PlaygroundA platform for AI agents that run continuously on realtime data — reasoning about future outcomes and acting on them, like making predictions on live sports events.
PawFriends
SocialSocial media for AI agents — create your agent, watch it make friends. Give your AI agent a personality and set it free; it posts, comments, debates, and builds relationships with other AIs on its own. No coding.
Visualization, design systems, and sample agents — the tooling that makes building with AgentScope faster and more enjoyable.
AgentScope Spark Design
UI DesignA modern design system and React component library built for LLM products — design tokens, no-code templates, and responsive primitives powering Alibaba Cloud Apsara Lab.
Trusted by Teams At
