MCP service access

Architecture

Capabilities

The following capability domains are covered. Select the corresponding API operations to expose them as MCP tools.

• The job diagnosis package recommended by Alibaba Cloud is used throughout this topic. It combines the diagnosis-related operations listed above. See Recommended APIs. For other packages such as job operations and data queries, select the corresponding operations in the same way. See API selection.

• Operations for AI Intelligent Assistant conversations (ChatAiAgent), intelligent inspection (ListPatrolReports, TriggerPatrol, and four more), and AI service management (GetFlinkAiService and others, sales-side foasconsole) are also available. See AI intelligent O&M.

Job diagnosis

Use case: locate the root cause of job failures, frequent failovers and restarts, startup failures, and performance degradation.

Example prompt: "One of my Flink jobs has been restarting frequently since this morning. Help me find out why."

The diagnosis chain is as follows.

The diagnosis result contains the following parts.

• Root cause category: back pressure, insufficient resources, code exception, or external dependency, taken from the diagnosis items returned by GetJobDiagnosis.

• Evidence: diagnosis item descriptions, key lines in the startup log, or the event timeline. Each item must be traceable to a tool response.

• Recommended action: use the fix suggestion returned with each diagnosis item, and check the recovery point and checkpoint health with ListSavepoints.

If the evidence is insufficient, ask the model to call ListSavepoints or ListAutopilotTuningHistories to restore the full context.

AI intelligent O&M

The AI Intelligent Assistant and intelligent inspection capabilities of Flink are available through OpenAPI. You can expose them as MCP tools in the same way and integrate them into your own agents and workflows. Both capabilities are also available in the console without any integration.

The management-side ververica operations cover both AI Intelligent Assistant conversations and intelligent inspection.

Flink AI service is the prerequisite for both capabilities. Its status, deletion protection, and free quota are exposed through the sales-side foasconsole operations.

The two paths are complementary: MCP exposure fits AI O&M capabilities into your own agents and observability stack, while the console entry ships with expert knowledge and built-in skills and works out of the box.

The AI Intelligent Assistant and intelligent inspection are billed by token usage, with a free trial quota before commercial launch. Quotas and billing details are shown on the AI Service page in the console.

Recommended APIs

The job diagnosis package recommended by Alibaba Cloud contains nine read-only operations. Together they cover the full chain of locating the job, checking instance status, diagnosing the root cause, and restoring the event timeline.

MCP instructions

MCP instructions guide the model to call tools in a defined workflow. The client must support the Instructions field of the MCP standard. Paste the following template into the MCP instructions field when you create the server.

You are a job diagnosis assistant for Realtime Compute for Apache Flink. When diagnosing a job issue, follow this workflow:
1) Locate the target job with ListDeployments or GetDeploymentsByName.
2) List job instances with ListJobs to identify the failed instance and the time of failure.
3) Call GetJobDiagnosis on the failed instance and read the diagnosis items, such as back pressure, insufficient resources, code exceptions, and network issues.
4) If the job failed to start, call GetLatestJobStartLog to read the exception stack traces in the startup log.
5) Restore the failover process with the event timeline from GetEvents. Refer to ListAutopilotTuningHistories when needed to determine whether Autopilot scaling is involved.
6) Output the conclusion: root cause category, evidence (quoting diagnosis items or key log lines), and recommended action.
This MCP server provides read-only capabilities only. It does not perform changes such as start and stop operations.

Procedure

1. Log on to the OpenAPI MCP Server console. In the left-side navigation pane, choose Custom > Create.

2. Configure the parameters as described in the following table.

3. Click Create and confirm the risk notice.

4. After the server is created, the page shows the Streamable HTTP Endpoint and the SSE Endpoint. Copy and store them. The endpoint uses the following format.

https://openapi-mcp.cn-hangzhou.aliyuncs.com/accounts/<account-id>/custom/<server-name>/id/<service-id>/mcp

OAuth

If your client supports remote MCP servers, add the server in the MCP settings of the client and set the URL to the Streamable HTTP Endpoint. Clients that require an explicit transport type accept streamable-http or sse. Qoder detects the transport type from the URL, so you do not need to specify it.

{
  "mcpServers": {
    "flink_diagnosis": {
      "type": "streamable-http",
      "url": "<your Streamable HTTP Endpoint>"
    }
  }
}

If your client supports only STDIO, such as Cursor or Lingma, bridge the connection with mcp-remote-alibaba-cloud.

{
  "mcpServers": {
    "flink_diagnosis": {
      "command": "npx",
      "args": [
        "mcp-remote-alibaba-cloud",
        "<your Streamable HTTP Endpoint>"
      ]
    }
  }
}

For Cursor, paste the configuration into ~/.cursor/mcp.json or .cursor/mcp.json in the project root. For Lingma, add the server on the MCP Tools page of the plug-in and select the STDIO type.

On the first connection, a browser window opens for authorization. Log on as a RAM user that belongs to the same Alibaba Cloud account as the MCP server and confirm the authorization. The client can then list and call the tools.

Static credentials

Use the local proxy alibabacloud.mcp-proxy with the AccessKey pair of a RAM user. Run the pre-check first to confirm that the proxy and the permissions are ready.

uvx alibabacloud.mcp-proxy@latest --server-url <MCP endpoint> pre-check

Add the MCP server in your client as an STDIO server with the following settings.

If you have signed in by running aliyun configure, the proxy reuses the credentials of the Alibaba Cloud CLI and you can omit the environment variables.