Develop an MCP service

Method 1: Deploy from a template

1. Log on to the Function AI console. On the Explore page, select the MCP Server template category, choose an MCP template, and then click its card. This topic uses the Amap template as an example.

2. On the template details page, click Deploy Now to open the service configuration page. Enter the API key that you obtained from the Amap Open Platform, and then click Deploy Project.

   The configuration page requires the AMAP_MAPS_API_KEY environment variable.

Method 2: Create from a blank project

1. Log on to the Function AI console. On the project page, choose Create Project > Create Blank Project, and then enter a custom name and description.

2. Click the blank project. On the project deployment page, choose New Service > MCP Service.

3. On the MCP service configuration page, configure the following options, and then click Preview & Deploy.

   • MCP service configuration

     Parameter	Description	Example
     Transport type	Supports hosting MCP services using the STDIO and SSE protocols. If your MCP service code uses the STDIO protocol, Function AI starts an SSE service to proxy STDIO requests. Clients must use SSE to access the service.	SSE
     Enable authentication	When you enable authentication, Function AI generates a read-only Bearer token that you can use to add built-in authentication to your service.	Turn on the switch
     Runtime	Use the official Node.js, Python, or Java runtimes to host your MCP service on the cloud with a single click and without code modification. To deploy with a container, you can also select custom container as the runtime. Then, follow the on-screen instructions to select an image hosted in ACR to start the build.	Node.js > Node.js 20
     Code repository	Use built-in code package You can use the Hello World runtime code package built into Function AI to simplify the deployment process. Later, you can write and debug your code in the WebIDE tab of the service details page. Bind a code repository for continuous deployment You can also bind a code repository to host your MCP service. This topic uses a GitHub repository as an example. After you bind a Git repository, any push to the specified branch automatically triggers continuous deployment for the service. Parameter descriptions: Build environment: For a multi-language project, you can select multiple build environments. Build command: npm build or pip install -t . -r requirements.txt Build cache: Optional. The cache directories for Node.js, Python, and Java runtimes can be set to ~/.npm, ~/.cache, and ~/.m2, respectively.	When deploying a Node.js project from a GitHub repository, select main for Branch, enter ./ for Code root directory, select Node.js 20 for Build environment, and enter npm install for Build command.

   • Resource configuration

     Parameter	Description	Example
     Instance type	Select or manually enter an appropriate combination of resource specifications based on your business requirements.	vCPU: 1 Memory: 2 GB
     Scaling policy	Default mode The most cost-effective mode. In this mode, Function Compute automatically allocates and releases instances for your function and can scale instances down to zero. This mode is suitable for scenarios with irregular and highly fluctuating traffic where some cold starts are acceptable. High-speed mode A balanced mode for cost and performance that effectively prevents cold starts. In this mode, you can lock elastic resources in advance by setting the number of provisioned snapshots for rapid scaling. You are charged for vCPU usage only while processing requests, which balances performance and cost. This mode is suitable for latency-sensitive scenarios. Note Because MCP SSE requests rely on a session mechanism, a request with the same session ID can lose context if routed to a different instance. Therefore, we recommend that you enable High-speed Mode, set one provisioned snapshot, and set the instance limit to 1. This ensures that SSE requests are preferentially routed to the same instance and reduces the probability of these requests being distributed to other elastic instances.	High-speed mode

   • Network configuration

     Parameter	Description	Example
     Allow VPC access	Specifies whether the function can access resources in a VPC. For more information, see Configure networking.	On
     VPC	Required when Allow VPC access is set to Yes. Create a new VPC or select the ID of the VPC that you want to access from the drop-down list.	fc.auto.create.vpc.1632317****
     vSwitch	Required when Allow VPC access is set to Yes. Create a new vSwitch or select a vSwitch ID from the drop-down list.	fc.auto.create.vswitch.vpc-bp1p8248****
     Security group	Required when Allow VPC access is set to Yes. Create a new security group or select a security group from the drop-down list.	fc.auto.create.SecurityGroup.vsw-bp15ftbbbbd****
     Disable public access URL	Turning on this switch restricts access to the MCP service to within the VPC, blocking it from the public internet.	On
     Allow specified VPCs only	To restrict calls to the MCP service to specific VPCs, turn on this switch and bind the desired VPCs.	On

Method 1: Test in the console

1. On the details page of your MCP service, go to the Service Test tab. Click Test Connection. After the connection succeeds, click Test Tool to start the test.

   If you enabled authentication (using Method 2), Function AI automatically enters the generated Bearer token during the test.

   In the Connection Information section of the Service Test page, click Test Connection. A success message is displayed. In the Debugging Information section, select the Tools tab. You can see the hello tool, with the description "Return string 'Hello World!'". Click Test Tool. A success message is displayed, and the service returns the result {"text": "Hello World!", "type": "text"}.

   If you use an incorrect Bearer token, the test fails.

2. After the test is successful, select the Monitoring and Logs tabs to view the test results.

Method 2: Test locally with Inspector

1. On your MCP service's details page, select the Access URL tab and get the public endpoint.

   The public endpoint format is similar to the following example: https://mcp-xxx.cn-hangzhou.fcapp.run/sse, where mcp-xxx is the application identifier.

2. Run the following command to start the Inspector tool locally:

   npx @modelcontextprotocol/inspector

3. In the URL input box of the Inspector tool, enter the public endpoint to test the connection.

   Set Transport type to SSE. After a successful connection, a green Connected status appears in the lower-left corner of the interface.

Create a custom MCP service

1. Log on to the Function AI console and find your MCP service. On its project details page, go to the Access URL tab and get the public endpoint.

2. Log on to the MCP management page of the Alibaba Cloud Model Studio console. Select the Custom Services tab, click + Create MCP Service in the upper-right corner, select Deploy with Script, and then click Deploy Service.

3. On the Create MCP Service page, set the Service Name, select http as the Installation method, fill in the MCP service configuration, and then click Submit Deployment.

   The following code provides an example of the MCP service configuration. Replace https://****.cn-hangzhou.fcapp.run/sse with the public endpoint that you obtained in Step 1.

   {
     "mcpServers": {
       "test-sse": {
         "url": "https://****.cn-hangzhou.fcapp.run/sse"
       }
     }
   }

Integrate with an Agent

1. Log on to the application management page of Alibaba Cloud Model Studio, choose + Create Application > Agent, and then click Create Now.

2. On the Agent page, select a model, add the MCP service that you created in the previous step, and then start testing.

   This topic uses the Qwen-Max model as an example. You can add multiple MCP services, including services that are pre-configured in Alibaba Cloud Model Studio and your custom MCP services.

   For example, after you add the fc-mcp service in the skills section, the UI shows that the service contains 12 tools.

Test the MCP service

• Large model application: Agent

• External tool: Custom location service tool get_geolocation

In the input box on the Text Chat interface, enter detailed location information for West Lake in Hangzhou and click the send button to start the test.

Configure service variables

On the Service Variables tab, add the following environment variables and enable the Encryption switch for each: AMAP_MAPS_API_KEY, AccessKeyID, and AccessKeySecret.

Load service variables

async function runServer() {
  console.log(`AMAP_MAPS_API_KEY: ${process.env.AMAP_MAPS_API_KEY}`)
  const transport = new StdioServerTransport();
  await server.connect(transport);
  console.error("Amap Maps MCP Server running on stdio");
}
runServer().catch((error) => {
  console.error("Fatal error running server:", error);
  process.exit(1);
});