Building Enterprise‑Grade AI Agents with JManus on Alibaba Cloud Serverless

JManus is a fully open‑source Java implementation of the OpenManus specification, part of the Spring AI Alibaba project, designed to let Java developers build enterprise‑grade, high‑availability AI agents quickly. It supports a multi‑agent framework, web‑based agent configuration, the Model Context Protocol (MCP), and the PLAN‑ACT execution model, and can integrate large models such as Claude 3.5 and Qwen 3.

Core Features

OpenManus multi‑agent support : Spring AI integration lets developers manage multiple agents with clear task division and collaboration.

MCP protocol : Agents can call local or cloud LLMs and interact with external services, APIs, and databases.

PLAN‑ACT mode : Enables complex reasoning, step‑by‑step execution, and dynamic adjustments for multi‑turn dialogs and automated workflows.

Performance Advantages

Compared with low‑code platforms like Dify, JManus running on Spring AI Alibaba delivers significantly better runtime performance and availability. Benchmarks show that the Serverless Java runtime (SAE) processes requests faster and with lower latency than Dify’s native agents.

Serverless Runtime – SAE & Function Compute

JManus leverages Alibaba Cloud’s Serverless Application Engine (SAE) for elastic, high‑availability deployment and Function Compute (FC) for secure, isolated tool execution. Key benefits include:

Isolation : Each agent runs in its own VPC and secure container, preventing resource contention.

Multi‑AZ deployment : Guarantees continuity across availability zones.

Canary and gray releases : SAE’s built‑in strategies enable risk‑free version rollouts.

Second‑level scaling : Automatic scaling based on CPU, memory, QPS, and response time.

Fast pod startup : SAE can launch pods in seconds, eliminating cold‑start delays.

Accelerated images : Alibaba DADI technology speeds up container image pull and start‑up.

JVM optimizations : JVM snapshot, class pre‑loading, and JIT tuning reduce Java start‑up time.

Long‑connection support : CLB/NLB gateways with keep‑alive keep SSE and other persistent connections stable.

Secure Sandbox for Tool Execution

Function Compute provides a millisecond‑level cold‑start and fine‑grained IAM‑based permission control. Each tool call runs in an isolated lightweight VM container, with read‑only file systems and disabled high‑risk operations, ensuring both security and resource efficiency.

MCP Marketplace Integration

FC can directly consume MCP services from the Alibaba MCP marketplace, allowing developers to add external tools (e.g., real‑time weather) with a simple JSON configuration.

{
  "mcpServers": {
    "openweather": {
      "type": "sse",
      "url": "https://example.fcapp.run/sse",
      "timeout": 45000
    }
  }
}

Deployment Steps

Obtain an OPENAI_API_KEY from Bailei.

In the SAE console, select the JManus template and configure VPC, API key, and optional public IP.

Confirm creation; SAE will provision the service and provide a public endpoint.

Optionally, deploy MCP services via Function Compute and add their URLs to the JManus configuration.

Create an agent in JManus, bind the MCP tool, and test (e.g., ask for real‑time weather).

The article includes several architecture diagrams illustrating the overall flow, SAE/FC deployment topology, and the JManus agent‑tool interaction.

Conclusion

By combining JManus with Alibaba Cloud’s Serverless SAE and Function Compute, developers obtain a powerful, secure, and highly scalable Java runtime for multi‑agent AI applications. The solution meets enterprise production requirements for high availability, elastic scaling, performance, and cost‑effective pay‑as‑you‑go billing, making it an ideal choice for cloud‑native AI agent development.