Designing Robust MCP Servers for Alibaba Cloud Observability 2.0 – Lessons & Best Practices

1. MCP Overview

MCP is an open protocol that standardizes how applications provide context to large language models (LLMs), similar to a USB‑C port for AI applications.

MCP Components

MCP Hosts : programs such as Claude Desktop, IDEs, or AI tools that want to access data via MCP.

MCP Clients : protocol clients that maintain a 1:1 connection with the server.

MCP Servers : lightweight programs that expose specific functions through the standardized model‑context protocol.

Local Data Sources : files, databases, and services accessible to the MCP server.

Remote Services : external systems reachable via APIs that MCP servers can connect to.

2. MCP Server for Alibaba Cloud Observability 2.0

Observability has evolved from simple monitoring to include logging, metrics, and distributed tracing, enabling teams to understand system behavior, locate issues quickly, and optimize performance. The new “UModel” provides a unified interaction language for multimodal data, supporting causal reasoning and proactive prediction.

2.1 Integration with Observability 2.0

By embedding MCP capabilities, users can converse with the observability platform in natural language, allowing the system to sense, analyze, and guide troubleshooting.

2.2 Design Practices

Key practices include simplifying tool interfaces, defaulting time parameters, limiting output size, avoiding chained tool calls, and using mock data for early testing.

3. Practical Experiences

3.1 Tool Interface Simplification

Tools should be understandable even by a three‑year‑old; complex SDK APIs like SLS get_logs must be wrapped into atomic, user‑friendly functions.

def get_log_tool(query: str, from_timestamp: int = Field(int(datetime.now().timestamp()) - 3600, description="from timestamp, unit is second"), to_timestamp: int = Field(int(datetime.now().timestamp()), description="to timestamp, unit is second")) -> list[Any]:

3.2 Default Time Parameters

Providing sensible defaults (e.g., last hour) covers most queries and reduces user errors.

3.3 Output Size Control

Limit JSON responses (e.g., top 10 items) to keep context manageable for LLMs.

3.4 Avoid Chain‑Calling Tools

Each tool should perform a single, independent task; chaining increases error risk.

3.5 Mock‑First Development

Define tool signatures and return mock data before implementing real APIs to ensure LLM compatibility.

4. Summary and Future Outlook

MCP standardizes LLM context interaction, supporting hosts, clients, servers, and data sources. Combined with Alibaba Cloud Observability 2.0 and UModel, it transforms passive monitoring into proactive, AI‑driven system insight. Future work will explore richer AI integrations and broader application scenarios.

References

https://github.com/CherryHQ/cherry-studio

https://deepchat.thinkinai.xyz/

https://help.aliyun.com/zh/model-studio/get-api-key

https://modelcontextprotocol.io/introduction

https://github.com/punkpeye/awesome-mcp-clients

https://modelcontextprotocol.io/clients

https://github.com/punkpeye/awesome-mcp-servers

https://modelcontextprotocol.io/examples

https://github.com/modelcontextprotocol/servers