What it does
AgentScope 2.0 is a Python agent framework that leans on a model's own reasoning and tool use rather than strict prompts and fixed orchestration. The SDK layer lets you compose an agent from building blocks: a ReAct reasoning-acting loop, a toolkit covering Python tools, MCP servers and skills with built-in shell, file-edit and search tools, model connectors, context management, memory with switchable backends, middleware, and permission controls with human-in-the-loop confirmation. It supports chat, embedding and text-to-speech models from providers including OpenAI, Anthropic, Gemini, DashScope, DeepSeek and Ollama, and tool execution can be isolated in sandboxes such as Docker, E2B, Daytona or Kubernetes. Recent additions listed on the page include a realtime voice agent, A2A protocol support and leader-and-member agent teams. A separate agent service wraps agents in a FastAPI backend with a pre-built web UI, multi-tenancy, scheduling, RAG and channels such as Feishu and Discord. AgentScope requires Python 3.11 or higher and is released under the Apache 2.0 license.
Capabilities
- ReAct agent loop with structured output and tool acting
- Toolkit for Python tools, MCP servers and skills
- Memory, context compaction and permission controls with human confirmation
- Sandboxed tool and code execution
- Leader-and-worker agent teams and task planning
- Agent service with a FastAPI backend and pre-built web UI
When to use it
- You want to build LLM agents in Python and wire in tools, MCP servers and memory
- You need multi-tenant, multi-session serving for agents behind a web UI
- You want agents to run code in an isolated sandbox
