AgentScope Runtime: A Production-grade Runtime for Agent Applications

[Cookbook] [Try WebUI] [中文README] [Samples]

Core capabilities:

Tool Sandboxing — tool call runs inside a hardened sandbox

Agent-as-a-Service (AaaS) APIs — expose agents as streaming, production-ready APIs

Scalable Deployment — deploy locally, on Kubernetes, or serverless for elastic scale

Plus

Full-stack observability (logs / traces)

Framework compatibility with mainstream agent frameworks

---

Table of Contents

Note

Recommended reading order:

• I want to run an agent app in 5 minutes: Quick Start (Agent App example) → verify with curl (SSE streaming)
• I care about secure tool execution / automation: Quick Start (Sandbox examples) → sandbox image registry/namespace/tag configuration → (optional) production-grade serverless sandbox deployment
• I want production deployment / expose APIs: Quick Start (Agent App example) → Quick Start (Deployment example) → Guides
• I want to contribute: Contributing → Contact

• News
• Key Features
• Quick Start: From installation to running a minimal Agent API service. Learn the three-stage AgentApp development pattern: init / query / shutdown.
  • Prerequisites: Required runtime environment and dependencies
  • Installation: Install from PyPI or from source
  • Agent App Example: How to build a streaming (SSE) Agent-as-a-Service API
  • Sandbox Example: How to safely execute Python/Shell/GUI/Browser/Filesystem/Mobile tools in an isolated sandbox
  • Deployment Example: Learn to deploy with DeployManager locally or in a serverless environment, and access the service via A2A, Response API, or the OpenAI SDK in compatible mode
• Guides: A tutorial site covering AgentScope Runtime concepts, architecture, APIs, and sample projects—helping you move from “it runs” to “scalable and maintainable”.
• Contact
• Contributing
• License
• Contributors

---

🆕 NEWS

• [2026-01] Added asynchronous sandbox implementations (BaseSandboxAsync, GuiSandboxAsync, BrowserSandboxAsync, FilesystemSandboxAsync, MobileSandboxAsync) enabling non-blocking, concurrent tool execution in async program. Improved run_ipython_cell and run_shell_command methods with enhanced concurrency and parallel execution capabilities for more efficient sandbox operations.
• [2025-12] We have released AgentScope Runtime v1.0, introducing a unified “Agent as API” white-box development experience, with enhanced multi-agent collaboration, state persistence, and cross-framework integration. This release also streamlines abstractions and modules to ensure consistency between development and production environments. Please refer to the CHANGELOG for full update details and migration guide.

---

✨ Key Features

• Deployment Infrastructure: Built-in services for agent state management, conversation history, long-term memory, and sandbox lifecycle control
• Framework-Agnostic: Not tied to any specific agent framework; seamlessly integrates with popular open-source and custom implementations
• Developer-Friendly: Offers AgentApp for easy deployment with powerful customization options
• Observability: Comprehensive tracking and monitoring of runtime operations
• Sandboxed Tool Execution: Isolated sandbox ensures safe tool execution without affecting the system
• Out-of-the-Box Tools & One-Click Adaptation: Rich set of ready-to-use tools, with adapters enabling quick integration into different frameworks

Note

About Framework-Agnostic: Currently, AgentScope Runtime supports the AgentScope framework. We plan to extend compatibility to more agent development frameworks in the future. This table shows the current version’s adapter support for different frameworks. The level of support for each functionality varies across frameworks:

Framework/Feature	Message/Event	Tool	Service
AgentScope	✅	✅	✅
LangGraph	✅	🚧	🚧
Microsoft Agent Framework	✅	✅	🚧
Agno	✅	✅	🚧
AutoGen	🚧	✅	🚧

---

🚀 Quick Start

Prerequisites

• Python 3.10 or higher
• pip or uv package manager

Installation

From PyPI:

# Install core dependencies
pip install agentscope-runtime

# Install extension
pip install "agentscope-runtime[ext]"

# Install preview version
pip install --pre agentscope-runtime

(Optional) From source:

# Pull the source code from GitHub
git clone -b main https://github.com/agentscope-ai/agentscope-runtime.git
cd agentscope-runtime

# Install core dependencies
pip install -e .

Agent App Example

This example demonstrates how to create an agent API server using agentscope ReActAgent and AgentApp. To run a minimal AgentScope Agent with AgentScope Runtime, you generally need to implement:

1. @agent_app.init – Initialize services/resources at startup
2. @agent_app.query(framework="agentscope") – Core logic for handling requests, must use stream_printing_messages to yield msg, last for streaming output
3. @agent_app.shutdown – Clean up services/resources on exit

import os

from agentscope.agent import ReActAgent
from agentscope.model import DashScopeChatModel
from agentscope.formatter import DashScopeChatFormatter
from agentscope.tool import Toolkit, execute_python_code
from agentscope.pipeline import stream_printing_messages

from agentscope_runtime.engine import AgentApp
from agentscope_runtime.engine.schemas.agent_schemas import AgentRequest
from agentscope_runtime.adapters.agentscope.memory import (
    AgentScopeSessionHistoryMemory,
)
from agentscope_runtime.engine.services.agent_state import (
    InMemoryStateService,
)
from agentscope_runtime.engine.services.session_history import (
    InMemorySessionHistoryService,
)

agent_app = AgentApp(
    app_name="Friday",
    app_description="A helpful assistant",
)


@agent_app.init
async def init_func(self):
    self.state_service = InMemoryStateService()
    self.session_service = InMemorySessionHistoryService()

    await self.state_service.start()
    await self.session_service.start()


@agent_app.shutdown
async def shutdown_func(self):
    await self.state_service.stop()
    await self.session_service.stop()


@agent_app.query(framework="agentscope")
async def query_func(
    self,
    msgs,
    request: AgentRequest = None,
    **kwargs,
):
    session_id = request.session_id
    user_id = request.user_id

    state = await self.state_service.export_state(
        session_id=session_id,
        user_id=user_id,
    )

    toolkit = Toolkit()
    toolkit.register_tool_function(execute_python_code)

    agent = ReActAgent(
        name="Friday",
        model=DashScopeChatModel(
            "qwen-turbo",
            api_key=os.getenv("DASHSCOPE_API_KEY"),
            stream=True,
        ),
        sys_prompt="You're a helpful assistant named Friday.",
        toolkit=toolkit,
        memory=AgentScopeSessionHistoryMemory(
            service=self.session_service,
            session_id=session_id,
            user_id=user_id,
        ),
        formatter=DashScopeChatFormatter(),
    )
    agent.set_console_output_enabled(enabled=False)

    if state:
        agent.load_state_dict(state)

    async for msg, last in stream_printing_messages(
        agents=[agent],
        coroutine_task=agent(msgs),
    ):
        yield msg, last

    state = agent.state_dict()

    await self.state_service.save_state(
        user_id=user_id,
        session_id=session_id,
        state=state,
    )


agent_app.run(host="127.0.0.1", port=8090)

The server will start and listen on: http://localhost:8090/process. You can send JSON input to the API using curl:

curl -N \
  -X POST "http://localhost:8090/process" \
  -H "Content-Type: application/json" \
  -d '{
    "input": [
      {
        "role": "user",
        "content": [
          { "type": "text", "text": "What is the capital of France?" }
        ]
      }
    ]
  }'

You’ll see output streamed in Server-Sent Events (SSE) format:

data: {"sequence_number":0,"object":"response","status":"created", ... }
data: {"sequence_number":1,"object":"response","status":"in_progress", ... }
data: {"sequence_number":2,"object":"message","status":"in_progress", ... }
data: {"sequence_number":3,"object":"content","status":"in_progress","text":"The" }
data: {"sequence_number":4,"object":"content","status":"in_progress","text":" capital of France is Paris." }
data: {"sequence_number":5,"object":"message","status":"completed","text":"The capital of France is Paris." }
data: {"sequence_number":6,"object":"response","status":"completed", ... }

Sandbox Example

These examples demonstrate how to create sandboxed environments and execute tools within them, with some examples featuring interactive frontend interfaces accessible via VNC (Virtual Network Computing):

Note

If you want to run the sandbox locally, the current version supports Docker (optionally with gVisor) or BoxLite as the backend, and you can switch the backend by setting the environment variable CONTAINER_DEPLOYMENT (supported values include docker / gvisor / boxlite etc.; default: docker).

For large-scale remote/production deployments, we recommend using Kubernetes (K8s), Function Compute (FC), or Alibaba Cloud Container Service for Kubernetes (ACK) as the backend. Please refer to this tutorial for more details.

Tip

AgentScope Runtime provides both synchronous and asynchronous versions for each sandbox type

Synchronous Class	Asynchronous Class
BaseSandbox	BaseSandboxAsync
GuiSandbox	GuiSandboxAsync
FilesystemSandbox	FilesystemSandboxAsync
BrowserSandbox	BrowserSandboxAsync
MobileSandbox	MobileSandboxAsync
TrainingSandbox	-
AgentbaySandbox	-

Base Sandbox

Use for running Python code or shell commands in an isolated environment.

# --- Synchronous version ---
from agentscope_runtime.sandbox import BaseSandbox

with BaseSandbox() as box:
    # By default, pulls `agentscope/runtime-sandbox-base:latest` from DockerHub
    print(box.list_tools()) # List all available tools
    print(box.run_ipython_cell(code="print('hi')"))  # Run Python code
    print(box.run_shell_command(command="echo hello"))  # Run shell command
    input("Press Enter to continue...")

# --- Asynchronous version ---
from agentscope_runtime.sandbox import BaseSandboxAsync

async with BaseSandboxAsync() as box:
    # Default image is `agentscope/runtime-sandbox-base:latest`
    print(await box.list_tools())  # List all available tools
    print(await box.run_ipython_cell(code="print('hi')"))  # Run Python code
    print(await box.run_shell_command(command="echo hello"))  # Run shell command
    input("Press Enter to continue...")

GUI Sandbox

Provides a virtual desktop environment for mouse, keyboard, and screen operations.

# --- Synchronous version ---
from agentscope_runtime.sandbox import GuiSandbox

with GuiSandbox() as box:
    # By default, pulls `agentscope/runtime-sandbox-gui:latest` from DockerHub
    print(box.list_tools())  # List all available tools
    print(box.desktop_url)  # Web desktop access URL
    print(box.computer_use(action="get_cursor_position"))  # Get mouse cursor position
    print(box.computer_use(action="get_screenshot"))  # Capture screenshot
    input("Press Enter to continue...")

# --- Asynchronous version ---
from agentscope_runtime.sandbox import GuiSandboxAsync

async with GuiSandboxAsync() as box:
    # Default image is `agentscope/runtime-sandbox-gui:latest`
    print(await box.list_tools())  # List all available tools
    print(box.desktop_url)  # Web desktop access URL
    print(await box.computer_use(action="get_cursor_position"))  # Get mouse cursor position
    print(await box.computer_use(action="get_screenshot"))  # Capture screenshot
    input("Press Enter to continue...")

Browser Sandbox

A GUI-based sandbox with browser operations inside an isolated sandbox.

# --- Synchronous version ---
from agentscope_runtime.sandbox import BrowserSandbox

with BrowserSandbox() as box:
    # By default, pulls `agentscope/runtime-sandbox-browser:latest` from DockerHub
    print(box.list_tools())  # List all available tools
    print(box.desktop_url)  # Web desktop access URL
    box.browser_navigate("https://www.google.com/")  # Open a webpage
    input("Press Enter to continue...")

# --- Asynchronous version ---
from agentscope_runtime.sandbox import BrowserSandboxAsync

async with BrowserSandboxAsync() as box:
    # Default image is `agentscope/runtime-sandbox-browser:latest`
    print(await box.list_tools())  # List all available tools
    print(box.desktop_url)  # Web desktop access URL
    await box.browser_navigate("https://www.google.com/")  # Open a webpage
    input("Press Enter to continue...")

Filesystem Sandbox

A GUI-based sandbox with file system operations such as creating, reading, and deleting files.

# --- Synchronous version ---
from agentscope_runtime.sandbox import FilesystemSandbox

with FilesystemSandbox() as box:
    # By default, pulls `agentscope/runtime-sandbox-filesystem:latest` from DockerHub
    print(box.list_tools())  # List all available tools
    print(box.desktop_url)  # Web desktop access URL
    box.create_directory("test")  # Create a directory
    input("Press Enter to continue...")

# --- Asynchronous version ---
from agentscope_runtime.sandbox import FilesystemSandboxAsync

async with FilesystemSandboxAsync() as box:
    # Default image is `agentscope/runtime-sandbox-filesystem:latest`
    print(await box.list_tools())  # List all available tools
    print(box.desktop_url)  # Web desktop access URL
    await box.create_directory("test")  # Create a directory
    input("Press Enter to continue...")

Mobile Sandbox

Provides a sandboxed Android emulator environment that allows executing various mobile operations, such as tapping, swiping, inputting text, and taking screenshots.

Prerequisites

• Linux Host: When running on a Linux host, this sandbox requires the binder and ashmem kernel modules to be loaded. If they are missing, execute the following commands on your host to install and load the required modules:

  # 1. Install extra kernel modules
  sudo apt update && sudo apt install -y linux-modules-extra-`uname -r`
  
  # 2. Load modules and create device nodes
  sudo modprobe binder_linux devices="binder,hwbinder,vndbinder"
  sudo modprobe ashmem_linux

• Architecture Compatibility: When running on an ARM64/aarch64 architecture (e.g., Apple M-series chips), you may encounter compatibility or performance issues. It is recommended to run on an x86_64 host.

# --- Synchronous version ---
from agentscope_runtime.sandbox import MobileSandbox

with MobileSandbox() as box:
    # By default, pulls 'agentscope/runtime-sandbox-mobile:latest' from DockerHub
    print(box.list_tools())  # List all available tools
    print(box.mobile_get_screen_resolution())  # Get the screen resolution
    print(box.mobile_tap([500, 1000]))  # Tap at coordinate (500, 1000)
    print(box.mobile_input_text("Hello from AgentScope!"))  # Input text
    print(box.mobile_key_event(3))  # HOME key event
    screenshot_result = box.mobile_get_screenshot()  # Get screenshot
    print(screenshot_result)
    input("Press Enter to continue...")

# --- Asynchronous version ---
from agentscope_runtime.sandbox import MobileSandboxAsync

async with MobileSandboxAsync() as box:
    # Default image is 'agentscope/runtime-sandbox-mobile:latest'
    print(await box.list_tools())  # List all available tools
    print(await box.mobile_get_screen_resolution())  # Get the screen resolution
    print(await box.mobile_tap([500, 1000]))  # Tap at coordinate (500, 1000)
    print(await box.mobile_input_text("Hello from AgentScope!"))  # Input text
    print(await box.mobile_key_event(3))  # HOME key event
    screenshot_result = await box.mobile_get_screenshot()  # Get screenshot
    print(screenshot_result)
    input("Press Enter to continue...")

Note

To add tools to the AgentScope Toolkit:

1. Wrap sandbox tool with sandbox_tool_adapter, so the AgentScope agent can call them:

   from agentscope_runtime.adapters.agentscope.tool import sandbox_tool_adapter
   
   wrapped_tool = sandbox_tool_adapter(sandbox.browser_navigate)

2. Register the tool with register_tool_function:

   toolkit = Toolkit()
   Toolkit.register_tool_function(wrapped_tool)

Configuring Sandbox Image Registry, Namespace, and Tag

1. Registry

If pulling images from DockerHub fails (for example, due to network restrictions), you can switch the image source to Alibaba Cloud Container Registry for faster access:

export RUNTIME_SANDBOX_REGISTRY="agentscope-registry.ap-southeast-1.cr.aliyuncs.com"

2. Namespace

A namespace is used to distinguish images of different teams or projects. You can customize the namespace via an environment variable:

export RUNTIME_SANDBOX_IMAGE_NAMESPACE="agentscope"

For example, here agentscope will be used as part of the image path.

3. Tag

An image tag specifies the version of the image, for example:

export RUNTIME_SANDBOX_IMAGE_TAG="preview"

Details:

• Default is latest, which means the image version matches the PyPI latest release.
• preview means the latest preview version built in sync with the GitHub main branch.
• You can also use a specified version number such as 20250909. You can check all available image versions at DockerHub.

4. Complete Image Path

The sandbox SDK will build the full image path based on the above environment variables:

<RUNTIME_SANDBOX_REGISTRY>/<RUNTIME_SANDBOX_IMAGE_NAMESPACE>/runtime-sandbox-base:<RUNTIME_SANDBOX_IMAGE_TAG>

Example:

agentscope-registry.ap-southeast-1.cr.aliyuncs.com/agentscope/runtime-sandbox-base:preview

Serverless Sandbox Deployment

AgentScope Runtime also supports serverless deployment, which is suitable for running sandboxes in a serverless environment, e.g. Alibaba Cloud Function Compute (FC).

First, please refer to the documentation to configure the serverless environment variables. Make CONTAINER_DEPLOYMENT to fc to enable serverless deployment.

Then, start a sandbox server, use the --config option to specify a serverless environment setup:

# This command will load the settings defined in the `custom.env` file
runtime-sandbox-server --config fc.env

After the server starts, you can access the sandbox server at baseurl http://localhost:8000 and invoke sandbox tools described above.

Deployment Example

The AgentApp exposes a deploy method that takes a DeployManager instance and deploys the agent.

• The service port is set as the parameter port when creating the LocalDeployManager.

• The service endpoint path is set as the parameter endpoint_path to /process when deploying the agent.

• The deployer will automatically add common agent protocols, such as A2A, Response API.

After deployment, users can access the service at http://localhost:8090/process:

from agentscope_runtime.engine.deployers import LocalDeployManager

# Create deployment manager
deployer = LocalDeployManager(
    host="0.0.0.0",
    port=8090,
)

# Deploy the app as a streaming service
deploy_result = await app.deploy(
    deployer=deployer,
    endpoint_path="/process"
)

After deployment, users can also access this service using the Response API of the OpenAI SDK:

from openai import OpenAI

client = OpenAI(base_url="http://0.0.0.0:8090/compatible-mode/v1")

response = client.responses.create(
  model="any_name",
  input="What is the weather in Beijing?"
)

print(response)

Besides, DeployManager also supports serverless deployments, such as deploying your agent app to ModelStudio.

import os
from agentscope_runtime.engine.deployers.modelstudio_deployer import (
    ModelstudioDeployManager,
    OSSConfig,
    ModelstudioConfig,
)

# Create deployment manager
deployer = ModelstudioDeployManager(
    oss_config=OSSConfig(
        access_key_id=os.environ.get("ALIBABA_CLOUD_ACCESS_KEY_ID"),
        access_key_secret=os.environ.get("ALIBABA_CLOUD_ACCESS_KEY_SECRET"),
    ),
    modelstudio_config=ModelstudioConfig(
        workspace_id=os.environ.get("MODELSTUDIO_WORKSPACE_ID"),
        access_key_id=os.environ.get("ALIBABA_CLOUD_ACCESS_KEY_ID"),
        access_key_secret=os.environ.get("ALIBABA_CLOUD_ACCESS_KEY_SECRET"),
        dashscope_api_key=os.environ.get("DASHSCOPE_API_KEY"),
    ),
)

# Deploy to ModelStudio
result = await app.deploy(
    deployer,
    deploy_name="agent-app-example",
    telemetry_enabled=True,
    requirements=["agentscope", "fastapi", "uvicorn"],
    environment={
        "PYTHONPATH": "/app",
        "DASHSCOPE_API_KEY": os.environ.get("DASHSCOPE_API_KEY"),
    },
)

For more advanced serverless deployment guides, please refer to the documentation.

---

📚 Guides

For a more detailed tutorial, please refer to:

---

💬 Contact

Welcome to join our community on

Discord	DingTalk

---

🤝 Contributing

We welcome contributions from the community! Here's how you can help:

🐛 Bug Reports

• Use GitHub Issues to report bugs
• Include detailed reproduction steps
• Provide system information and logs

💡 Feature Requests

• Discuss new ideas in GitHub Discussions
• Follow the feature request template
• Consider implementation feasibility

🔧 Code Contributions

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

For detailed contributing guidelines, please see CONTRIBUTE.

---

📄 License

AgentScope Runtime is released under the Apache License 2.0.

Copyright 2025 Tongyi Lab

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

✨ Contributors

Thanks goes to these wonderful people (emoji key):

Weirui Kuang 💻 👀 🚧 📆	Bruce Luo 💻 👀 💡	Zhicheng Zhang 💻 👀 📖	ericczq 💻 📖	qbc 👀	Ran Chen 💻	jinliyl 💻 📖
Osier-Yi 💻 📖	Kevin Lin 💻	DavdGao 👀	FlyLeaf 💻 📖	jinghuan-Chen 💻	Yuxuan Wu 💻 📖	Fear1es5 🐛
zhiyong 💻 🐛	jooojo 💻 🐛	Zheng Dayu 💻 🐛	quanyu 💻	Grace Wu 💻 📖	LiangQuan 💻	ls 💻 🎨
iSample 💻 📖	XiuShenAl 💻 📖	Farruh Kushnazarov 📖	fengxsong 🐛	Wang 💻 🐛	qiacheng7 💻 📖	Yuexiang XIE 👀
RTsama 🐛 💻	YuYan 📖	Li Peng (Yuan Yi) 💻 📖 💡	dorianzheng 👀 📦
Add your contributions

This project follows the all-contributors specification. Contributions of any kind welcome!