Publish and manage MCP tools

Add a tool

1. Select the application and application version, specify the Tool Name, and select the Owning MCP server.

   

   Parameter	Description
   Application and application version	Important Relationship between tools and RPA applications: You can only create an MCP Tool from a published application. There is a one-to-one relationship between an application and a tool, as well as between an application version and a tool version.
   Tool Name	The Tool Name can be up to 100 characters long. Chinese characters are not supported. We recommend using underscores (_) to separate words. Important We recommend not using the application name as the Tool Name. In on-premises robot scenarios, users can also filter and execute applications by name in the robot client. The large model uses the Tool Name and Tool Description to find the tool, so the name should be easy for the model to understand.
   Owning MCP server	You can only select a business MCP Server.

2. Add the tool description and define the input parameters for this tool version.

   

   Parameter	Description
   Description (Tool description)	The product provides a default Description template that you can customize.
   Input parameters	Input parameters are divided into two types: input parameters for the automated workflow and system input parameters. System input parameters cannot be modified or deleted. For automated workflow input parameters, the parameter name, parameter type, and required status are defined during the workflow development phase. Important This step validates the parameter name and type. If validation fails, you cannot proceed. Follow the error message to make corrections in the editor's parameter panel, republish the application, and then add the tool for the new version. The parameter name, type, and required status must be defined or modified in the editor's parameter panel, as shown in the following figure.

   • Input parameters for the automated workflow

     Item	Details
     Parameter name	Define and edit in the editor's parameter panel. Format requirements: Must consist of uppercase or lowercase letters, numbers, or underscores. It must start with a letter and must not end with an underscore.
     Parameter type	Editor's parameter panel Note When you develop an MCP Tool by using RPA, use the "text box" type for all parameters, as shown in the preceding figure. Perform format conversions within the automated workflow as needed. This approach prevents data transmission issues caused by format conversion. Tool parameter type The MCP protocol uses JSON-RPC to describe tool definitions. Therefore, when you define tool input parameters in this step, use only basic JSON types and JSON Schema. For recommendations on using specific parameter types, see the following sections.
     Required	Define and edit in the editor's parameter panel.
     Parameter description	You must provide a detailed description to inform the large model about the purpose of this input parameter and any additional requirements for its content, such as format rules.

     Parameters can be classified into three types based on their purpose: basic types, structured strings, and files.

     Basic types

     Supported basic types include string, integer, date, and datetime. You can also set validation conditions for basic types, such as default values and enum values.

     Structured strings

     Automated workflows often need to operate on a set of data, such as sending different messages to multiple people. In these cases, an input parameter may require a specific data structure. When you develop an MCP Tool by using RPA, keep the structure of input parameters simple. Complex schemas make it difficult for the large model to generate tool inputs.

     For structured input parameters, you can select the "JSON structure" type and follow the guide to complete the definition.

     The complexity of the parameter structure should not exceed that of the following examples:

     Type	Example
     Simple array	["1","2","3"]
     Two-dimensional array	[ ["1","2","3"], ["a","b","c"] ]
     Key-value structure	{ "name":"Tom", "age":21, "address":"Beijing" }
     Array of key-value structures	[ { "name": "Tom", "age": 21, "address": "Beijing" }, { "name": "Jerry", "age": 19, "address": "Hangzhou" } ]

     You can use the built-in feature to automatically generate the parameter definition.

     • Select "JSON structure" for the parameter type to open the JSON definition window. If you are familiar with JSON Schema, you can define the data structure directly in the window or import existing JSON Schema content.

     • 

     • You can also use the "Quick JSON Data Import" feature to automatically generate the corresponding JSON schema from JSON data.

     Note

     • We recommend defining structured input parameters to help the large model understand and generate the required tool parameters.

     • We recommend using the "Quick JSON Data Import" feature to automatically generate the definition from sample data.

     Files

     Automated workflows often need to handle files, such as reading data from an Excel file to fill out a web form. In such cases, you need to consider how to pass file content to the automated workflow.

     • Method 1: If the file content is small and has a clear structure (such as Excel or CSV), you can parse the content and pass it as a structured input parameter as described earlier.

     • Method 2: If the file is large, we recommend passing a file link. The input parameter for the automated workflow is a file link (for example, a file stored in Object Storage Service (OSS)). The workflow then downloads the file, parses it, and performs the subsequent operations. For more information, see Download and upload files by using Alibaba Cloud OSS.

   • System input parameters

     Parameter	Type	Required	Description
     RobotId	string	Yes	Specifies the robot that executes the automation task. Both on-premises robots and service robots can be invoked by MCP.
     TaskPriority	integer	No	The priority of the automation task. A positive integer from 0 to 100. The default value is 50. This setting affects only the task's position in the robot's pending execution queue. Higher values indicate higher priority.

3. (Optional) Configure callback settings.

   1. After you configure a callback, a notification is triggered when the status of a task created by this MCP Tool changes. The details of the callback notification depend on the robot type used. For a service robot, see CreateTask. For an on-premises robot, see createServiceTask.

   2. When you configure a callback URL, an HTTP POST request is sent to the URL to verify connectivity.

      Important

      Validation involves sending an empty POST request to the URL over the internet. The test passes if the server returns an HTTP 200 status code.

4. (Optional) You must publish a tool to an MCP gateway instance before it can be invoked. You can publish the newly defined tool directly in this step.

5. (Optional) Use a large model to evaluate the MCP Tool. This helps you refine the information you have added to the tool.

Evaluate a tool

Before an MCP Tool is invoked by a large model, you can use the AI-assisted evaluation feature. From the model's perspective, this feature evaluates whether the tool's information can be accurately understood and used. You can then decide whether to make changes based on the evaluation results.

• The evaluation covers the Tool name, description, and input schema for this version of the MCP Tool.

• The evaluation process is static and does not trigger an actual MCP Tool invocation.

• The following figure shows an example of evaluation results:

Tool version management

Testing and production for iterations

During the continuous development and iteration process after a tool is published, you may need to make the following types of changes:

• Modifications to the tool's information, such as changing the description.

• Adjustments to the automated workflow, which result in a new version of the corresponding RPA application.

In this case, you also need to test the tool before going live. Similar to the process for RPA applications (for more information, see Application iteration: testing and release), you can create a separate MCP Server and MCP Tool for testing purposes. Take get_city_weather as an example.

Other operations

• Edit a specific tool version. You can modify the tool's description and input parameters. After modification, you must publish the tool to the MCP gateway instance for the changes to take effect.

• Edit the Tool Name. You must unpublish all versions of the tool before you can modify its name.

• Delete a tool version. You must unpublish the version before you can delete it.

• Delete a tool. After you delete all versions of a tool, the tool is automatically deleted.

System built-in tools

For more information, see System built-in MCP Tools.

MCP invocation authentication and permissions

All MCP requests require authentication. Anonymous invocations are not supported. You can manage credentials in MCP Server > Credential Management.

The permission scope for MCP invocation depends on the robot type used. For more information about robot types, see Choose a robot type.

• On-premises robot

  • A robot can run only the applications that have been approved for its account in advance. For more information, see Publish and manage applications. The robot can be triggered by manual execution in the client or an OpenAPI call, among other methods.

  • If you use MCP to invoke this type of robot, the MCP Client still receives the list of all tools on the MCP gateway instance. However, tool execution may fail if the token lacks the required application permissions.

• Service robot

  • A service robot can be triggered by an OpenAPI call. OpenAPI can run all published applications; your upstream system is responsible for controlling the execution scope.

  • Therefore, when you use MCP to invoke this type of robot, the MCP Client can invoke all available tools.

MCP connection information

In the MCP Server menu, you can select "Get MCP Server Connection Information" to obtain the invocation address and connection details for a specific MCP gateway instance.

MCP invocation sequence

See OpenAPI triggers and MCP triggers.