Custom agent deployment

Source file type

Source file type: Choose the source for your application: Git repository, OSS file, or container image. The parameters that follow change based on this selection.

Source type availability varies by deployment edition:

Container image

The image is automatically pulled to the ECS instance during deployment before the script is executed. You do not need to run docker pull in the script.

We recommend building your image based on the official public ADK base image.

FROM compute-nest-registry.cn-hangzhou.cr.aliyuncs.com/computenest/agentscope-adk-base-py312:latest
WORKDIR /app
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY . .
CMD ["bash", "docker-entrypoint.sh"]   # Start

Git repository (ECS edition only)

The code is automatically downloaded to code_deploy_application/. You do not need to run git clone.

OSS file (ECS edition only)

The file is automatically downloaded to the working directory. If it is a compressed archive, you must unzip it in the application startup script.

Application connections

Session connections persist the context of multi-turn conversations and correspond to Redis instances. To use them, you must first create an instance in the Redis Console and then create a connection in Agent Deployment > Connection Management. For more information, see Connection Usage Instructions.

ECS edition: Optional. If you create a new VPC, the list does not display session connections, and you can use the built-in local Redis. If you select an existing VPC, the list displays only connections that are in the same network as the VPC, and you can bind them as needed.

Container cluster edition: Required. You must select an existing VPC and bind a session management connection that is in the same network as the VPC.

For the ECS edition, session connection options are not displayed when you opt to create a new VPC. This is because the VPC does not exist at configuration time, preventing the platform from matching it to the Redis network.

MCP configuration

Example MCP JSON format:

{
  "mcpServers": {
    "WebSearch": {
      "type": "streamableHttp",
      "baseUrl": "https://dashscope.aliyuncs.com/api/v1/mcps/WebSearch/mcp",
      "headers": {
        "Authorization": "Bearer ${DASHSCOPE_API_KEY}"
      }
    }
  }
}

${DASHSCOPE_API_KEY} is a placeholder that is replaced with the actual API Key by the platform during deployment.

Example code for reading the configuration:

The ECS edition reads from a file. The container cluster edition reads from MCP_JSON and decodes the content by using Base64.

import base64
import json
import os
from pathlib import Path

def load_mcp_config() -> dict:
    if raw := os.getenv("MCP_JSON"):
        return json.loads(base64.b64decode(raw).decode("utf-8"))
    path = Path(os.getenv("MCP_CONFIG_PATH", "/root/mcp_config.json"))
    return json.loads(path.read_text(encoding="utf-8"))

During debugging, you can run echo "$MCP_JSON" | base64 -d to view the decoded JSON.

The baseUrl for each MCP is subject to the value displayed in the Model Studio (Bailian) console. You must enable the corresponding MCP capability in Model Studio (Bailian) before use. For more information, see Connection Guide.

Skill configuration

The code can read installed Skill files from /root/skills for the Agent to load on demand.