How to Automate HubSpot CRM Using OpenAI Agent Builder

OpenAI's .

, which describes building and deploying AI Agents from scratch. You can also play with the templates provided by OpenAI to get started with the Agent Builder.

They also provide a comprehensive Evaluation Tool to test the performance of your Agent and identify bottlenecks, so you don't need anything further to test your workflows.

existing toolkit infrastructure. Instead of writing custom OAuth flows for each MCP or managing API keys, you can use Rube as a unified MCP interface for more than 500 apps, including HubSpot, Salesforce, Google Sheets, Airtable, Notion, etc., which work seamlessly with the Agent Builder and most other MCP Clients.

You can ask Rube to authenticate you with the app (if your app works on OAuth2); it'll handle the OAuth flow for you and get the access token. If your app works on an API key, you can just provide the API key to Rube, and it'll use it to make the API calls for you.

How to Add HubSpot to Agent Builder

Before we start building our CRM agent, we need to have a few things in place:

Step 1: Set up Rube MCP

1. Go to
   
   
   
   
   
   1. In the modal, select Agent Builder. Copy the MCP URL. We’ll paste this into Agent Builder soon.
   
   
   2. Scroll a little further and hit the "Generate Token" button to generate the access token.
   
   
   
   
   
   
   
   
   
   1. 
      

      Click the "Agent" node, name the agent "HubSpot CRM Agent," and give the agent instructions as below:
      
      

      
      
      CODE Kopieren

      You are a CRM agent that can help me manage my contacts and deals in HubSpot CRM.
      
      

      
      
      
      

      
      

      Note: We'll be editing this node later with the inputs for adding guardrails and logical nodes to the agent.

      
      

      
      
   
   

   2. Click on the "+" icon beside Tools in the agent node's configuration panel. It will open a dropdown; select "MCP Server" → click the "+ Server" button in the dialogue that appears.

   
   

   3. Paste the MCP URL you copied from Rube, and in the Authorisation field, choose Access token / API Key and paste the token you generated.

   
   
   
   
   
   
   

   
   
   
   Step 2: Add Intent Classification
   

   
   
   
   
   
   • Now, create a Classification Agent for the workflow to classify the user's intent and route the request to the appropriate node. For this guide, we'll focus on creating a simple classification agent to classify the user's intent to create, update, or delete a contact and route the request to the appropriate node.
   
   
   • 
     

     Drag and drop the Agent node from the toolbar, connect it with the "Pass" option of the Guardrails node, and in the configuration panel, give the Agent a label of "Classification Agent." Give the agent instructions as below:
     
     

     
     
     CODE Kopieren

     Classify the user’s intent into one of the following categories: "create_contact", "update_contact", or "delete_contact".
     
     

     
     
     
   
   
   • 
     

     For the output format, select "JSON," and in the JSON schema field, add the following schema by going to Advanced Settings:
     
     

     
     
     CODE Kopieren

     {
     
       "type": "object",
     
       "properties": {
     
         "classification": {
     
           "type": "string",
     
           "enum": [
     
             "create_contact",
     
             "remove_contact",
     
             "update_contact"
     
           ],
     
           "description": "classification of user intent",
     
           "default": ""
     
         }
     
       },
     
       "additionalProperties": false,
     
       "required": [
     
         "classification"
     
       ],
     
       "title": "response_schema"
     
     }
     
     

     
     
     
     

     Update the schema by clicking the "Update" button in Advanced Settings. For the "Fail" option, we will use the "End" node to terminate the workflow, ensuring that if the GuardRails node fails, the workflow will conclude.

     
     
   
   
   
   
   

   
   
   
   Step 3: Route Logic
   

   
   
   
   
   
   • 
     

     Now, we need to use the output variables from the Classification Agent in the next node, which will be a logical node to route the request to the appropriate node based on the user's intent.

     
     
     
     
     
     • 
       

       Drag and drop the "If/else" node from the toolbar, connect it with the "Classification Agent" node, and in the configuration panel, give a CaseName to the logic: "isValid." Then pass this expression in the next input field:
       
       

       
       
       CODE Kopieren

       input.output_parsed.classification == "create_contact" || input.output_parsed.classification == "update_contact" || input.output_parsed.classification == "delete_contact"
       
       

       
       
       
     
     

     • For the "Else" option, connect it with the "End" node to end the workflow.

     
     
     
     
     
     
   
   
   
   
   
   
   
   

   For deploying the agent, you can click the "Publish" button, give a name to your workflow, and click on the "Publish" button to deploy the workflow.

   
   
   

   , , Voice Agents (speech-to-speech architecture)

   
   
   

   
   
   
   Conclusion
   

   
   
   

   In this guide, we built a HubSpot CRM Agent using OpenAI’s Agent Builder, integrating it with Rube MCP and adding guardrails, logical nodes, and intent classification to make the workflow secure and efficient. In practice, when I was working with the MCP node, it became quite complex to parse the desired payload requests from the agents across MCPs for their schemas. Transform nodes can help stabilize the input, but using them can also make the workflow significantly more complex.

   
   
   

   Nevertheless, Agent Builders are powerful because they let you design complex AI workflows visually, combine multiple tools and APIs, and focus on the logic of your agents rather than the underlying infrastructure. They make it easier to automate repetitive tasks, connect with various services, and quickly iterate on your agent’s behavior without writing a lot of boilerplate code.

   Vollständiger Original-Bericht

   Ausführliche Details, Code-Beispiele & Hersteller-Stellungnahme auf dev.to.

   ↗ Original-Artikel auf dev.to lesen