Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

agentscope-ai/agentscope

AI Engineering

Python framework for building multi-agent apps with tools, memory, sandboxes and a deployable agent service.

36,301
3,891
Python
Apache-2.0
2 years old

What it is

AgentScope 2.0 is a Python framework for building LLM agents that rely on the model's reasoning and tool-use abilities rather than rigid prompt chains. It provides the building blocks for a single agent (ReAct loop, toolkit, model adapters, context management, event streaming, permissions, middleware, memory, workspace/sandbox) and ships a batteries-included agent service for turning agents into a multi-tenant, multi-session application with a FastAPI backend and a pre-built Web UI. It targets developers who need to prototype an agent quickly and later deploy it as a service; providers include OpenAI, Anthropic, Gemini, DashScope, DeepSeek, Moonshot, Volcengine, xAI and Ollama, and tools can come from Python functions, MCP servers or skills.

Recent additions include SOP-style multi-step tasks, a leader–worker TeamPipeline, realtime voice agents over WebRTC, voice cloning, model routing middleware and IM channel integrations such as Feishu and Discord. Check the documentation for the full conceptual model.

Quick start

Requirements: Python 3.11 or higher. Install from PyPI:

uv pip install agentscope

Alternatively, install from source in editable mode:

git clone -b main https://github.com/agentscope-ai/agentscope.git
cd agentscope
uv pip install -e .

Set DASHSCOPE_API_KEY and run a minimal agent in the terminal:

from agentscope.agent import Agent
from agentscope.console import launch_console
from agentscope.tool import Toolkit, Bash, Grep, Glob, Read, Write, Edit
from agentscope.credential import DashScopeCredential
from agentscope.model import DashScopeChatModel

import os, asyncio

async def main() -> None:
    agent = Agent(
        name="Friday",
        system_prompt="You're a helpful assistant named Friday.",
        model=DashScopeChatModel(
            credential=DashScopeCredential(api_key=os.environ["DASHSCOPE_API_KEY"]),
            model="qwen3.6-plus",
        ),
        toolkit=Toolkit(tools=[Bash(), Grep(), Glob(), Read(), Write(), Edit()]),
    )
    await launch_console(agent)

asyncio.run(main())

launch_console handles streamed output, tool-call confirmation and Ctrl+C interruption. The example uses DashScope and its qwen3.6-plus model; you can swap the model adapter for another provider.

To run the agent service, start the backend from examples/agent_service with python main.py, then start the Web UI from examples/web_ui with pnpm install and pnpm dev. The service backend needs model credentials configured in its own setup.

Constraints and operational notes

The framework requires Python 3.11+. The agent service is a FastAPI backend plus a Node/pnpm front-end, so running the full stack requires both Python and Node tooling. Tool execution can be isolated across local, Docker, Apple Container, Bubblewrap, E2B, OpenSandbox, Daytona and Kubernetes sandboxes; which ones are available depends on your environment. Permission and human-in-the-loop controls allow tool-call confirmation or a bypass mode, which should be chosen deliberately because bypass mode lets the agent run tools without pausing. Memory backends are switchable (ReMe, Mem0). The project is released under Apache-2.0, and its research is documented in the AgentScope 1.0 paper and the earlier multi-agent platform paper.

HotRepo's take

On 2026-11-19 the Python daily list recorded AgentScope at 36,293 stars and 3,891 forks, with 75 stars gained that day and a best position of 21 on that language list. That is a single daily snapshot rather than a long trend, so it supports noting sustained attention on the Python side but not a longer growth claim. The notable technical signal in the snapshot period is scope: the project combines a low-level agent SDK with a deployable multi-tenant service, realtime voice and team orchestration in one codebase, which positions it for developers who want to move from a scripted prototype to a hosted agent application without switching frameworks.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

Hacker News

AgentScope
▲ 1 · 0 comments · September 10, 2025

You might also like

  1. agentscope-ai/agentscope-javaA Java framework for running long-lived, distributed LLM agents with permissions, event streams and persistent memory.Java6,136
  2. bojieli/ai-agent-bookA free, open-source book with source text, PDF/EPUB downloads, and 109 hands-on AI Agent experiments for developers.Python58,639
  3. AstrBotDevs/AstrBotA self-hostable bot framework that puts an LLM agent into your existing chat platforms, with plugins and a sandbox.Python43,337
  4. pinganbbOvO/ai-companion-kitA toolkit for building a QQ chatbot with persistent memory and an MCP server that controls an Android phone.JavaScript572
  5. QwenAudio/qwen-audio-agentA realtime voice runtime that lets an AI agent keep talking while it hands tasks to a backend coding agent.JavaScript3,078