agentscope-ai / agentscope

**中文主页** | **Documentation** | **Roadmap**
## What is AgentScope 2.0?
AgentScope 2.0 is a production-ready, easy-to-use agent framework with essential abstractions that keep up with rising model capability.
We design for increasingly agentic LLMs.
Our approach leverages the models' reasoning and tool use abilities
rather than constraining them with strict prompts and opinionated orchestrations.
## News
- **[2026-10] `FEAT` `Experimental`:** Agent service supports realtime voice conversations over WebRTC, with text and voice sharing the same session history.
- **[2026-09] `FEAT` `Experimental`:** SOP supported — run multi-step tasks as standard operating procedures, in both the framework and agent service. Docs | Service
- **[2026-09] `FEAT`:** `TeamPipeline` supported — a leader agent delegates tasks to member agents. Docs
- **[2026-09] `INTE`:** MiniMax chat models supported via `MiniMaxChatModel` (MiniMax-M3 by default).
- **[2026-09] `FEAT`:** Model routing supported — use `ModelRouterMiddleware` to select an appropriate chat model for each reply.
- **[2026-09] `INTE`:** Jev classifier model supported through the TypeSafe SDK via `JevClassifierModel`.
- **[2026-09] `FEAT`:** agentscope-skill now supports AgentScope v2.
- **[2026-09] `INTE`:** Support DashScope, OpenAI, Gemini and xAI realtime APIs in `RealtimeAgent`. Example | Docs
- **[2026-09] `FEAT` `Experimental`:** Realtime voice agent supported. Example | Docs
- **[2026-09] `FEAT`:** A2A protocol supported — chat with any remote A2A agent via `A2AAgent`. Example | Docs
More news →
## Community
Welcome to join our community on
| Discord | DingTalk |
|----------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------|
| | |
## Quickstart
### Installation
> AgentScope requires **Python 3.11** or higher.
#### From PyPI
 [code] 
#### From source
 [code] 
## Agent
The SDK layer — compose an agent from a rich set of building blocks:
| Building block | What's inside |
|---|---|
| **ReAct** | Reasoning-acting loop with structured output, realtime interruption & resume, and batched (sequential / concurrent) tool acting |
| **Toolkit** | Agentic tool management over Python tools, MCP servers, and skills; ships with built-in coding tools (shell, file edit, search) and task/plan tools |
| **Model** | LLM, embedding, and TTS across major providers (OpenAI, Anthropic, Gemini, DashScope, DeepSeek, Moonshot, Volcengine, xAI, Ollama) |
| **Context** | Automatic compaction, tool-result offload, and context injection (system prompt, RAG, memory) via built-in middleware |
| **Event System** | Unified event bus streaming reasoning, tool calls, and multimodal content (text, image, audio) to the frontend |
| **Permission & HITL** | Fine-grained control over tools and resources, confirmation, bypass mode |
| **Middleware** | Composable hooks across the loop — reply, reasonin