Ever been asked to run AI inside HubSpot, but on your model? Maybe legal wants the data to stay in your Azure tenant. Maybe you self-host a fine-tuned model and refuse to send prompts anywhere else. Maybe you just found a cheaper, faster model on OpenRouter and want it in your workflows today.

Until now, HubSpot said no. Its native custom LLM action is Enterprise-only, it's limited to a short list of named vendors, and there's no place to type a base URL of your own. So you were stuck picking from someone else's menu.

Not anymore. Say hello to Send to Custom AI, the newest action in AI Studio for HubSpot.

What it does

Send to Custom AI runs a prompt on any OpenAI-compatible endpoint you control, straight from a HubSpot workflow. If your model answers the standard chat completions call over HTTPS, it works. That covers a lot of ground:

  • Self-hosted models on vLLM, Ollama, or LM Studio, behind a TLS tunnel.
  • OpenRouter, so one key gives you hundreds of models to route between.
  • Azure OpenAI on the v1 endpoint, so inference stays in your tenant and your region.
  • Groq, Together, Fireworks, DeepInfra, Mistral, and other hosted APIs.
  • Gateways like LiteLLM or Portkey that you already run for routing and spend control.

The model id is just free text. So the day a new model ships, you type its name and you're off. No waiting on a vendor list.

Your key never touches HubSpot

Here is the part I care about most. HubSpot workflow fields have no secret type, which means anyone who can edit a workflow can read what's in them. Pasting an API key there? Bad idea.

So we don't. Instead, you save your endpoint once as an AI Connector at ai.resonatehq.com: a name, a base URL, an API key, and a default model. The key is encrypted with AES-256-GCM, stored write-only, and decrypted only when a run actually fires. In the workflow, you just pick the connector from a masked dropdown. The key is never entered in HubSpot, never shown in a workflow, and never copied when someone clones that workflow.

And there's no third-party gateway in the middle. Your prompt goes straight to the endpoint you chose, over TLS, secured by your own key. We don't route it through some model middleman, and by default we keep only run metadata: the connector, host, model, and status. Prompt and response logging stays off unless you turn it on per connector.

How it looks in a workflow

Here is a real example. Enrich a company, then have your own hosted model classify the account tier and write it back to the record.

Trigger enrollment for companies
When this happens
Group 1
Company enrichment status is Needs review any number of times anytime
+
1. Send to Custom AI

Classify the account tier on your own hosted model

Click to see configuration
+
2. Edit record

Set Account tier to the outcome of Send to Custom AI

+
End

Notice the Advanced setup toggle. By default you see the essentials: connector, model, prompt. Flip it on when you want to tune temperature, reasoning effort, max tokens, or pass extra JSON parameters. Most workflows never need it, so we keep it out of your way.

Test before you trust it

Nobody likes shipping a workflow and finding out at 2am that the endpoint was misconfigured. So AI Studio lets you test a connector live before you save it, and rehearse full prompts in the Playground. Both run the exact same request the workflow will, so a green check really does mean production will work. And both are unbilled.

How it works

  1. Add an AI Connector in AI Studio with your base URL, key, and default model.
  2. Hit Test connection. Green means go.
  3. Drop Send to Custom AI into any workflow and pick the connector.
  4. Use the output. Save it to a property or note, split it into fields, or branch on the status.

Every run returns a clear status, so your workflow can retry, route, or notify on a failure instead of writing error prose into a CRM property. That alone is worth the switch.

What it costs

You pay your own provider for the inference, on your own key. AI Studio charges a flat 50 credits per successful run. Failed runs cost nothing, and Playground tests are unbilled. Simple.

One more thing: Breeze agents

The same connectors also power an Ask Custom AI tool for HubSpot's Breeze agents, right alongside Ask AI and Search the Web. Same encryption, same billing, same endpoints. (HubSpot agent tools are in beta for now.)

So whether you run on Azure for compliance, OpenRouter for cost, or a model you fine-tuned yourself, your CRM can finally use it. On your terms, with your key, inside the HubSpot workflows your team already lives in.

Ready to bring your own model? Add a connector and point a workflow at it. Or read the full Send to Custom AI guide first.

Which endpoints work with Send to Custom AI?+
Any OpenAI-compatible endpoint that answers POST {base URL}/chat/completions over HTTPS. That covers self-hosted vLLM and Ollama, OpenRouter, Azure OpenAI on the v1 endpoint, Groq, Together, Fireworks, DeepInfra, Mistral, and gateways like LiteLLM and Portkey. The model id is free text, so you are not limited to a fixed vendor list.
Where does my API key live?+
In an AI Connector at ai.resonatehq.com, not in HubSpot. Your key is encrypted with AES-256-GCM, stored write-only, and the workflow only sees a masked dropdown. HubSpot workflow fields have no secret type, so keys never belong there.
Do I need HubSpot Enterprise?+
No. HubSpot's native custom LLM action is Enterprise-only and limited to named vendors with no custom base URL. Send to Custom AI works on the HubSpot tiers AI Studio supports and takes any endpoint you control.
What does it cost?+
A flat 50 credits per successful run. You pay your own provider for the inference. Failed runs cost nothing, and Playground tests are unbilled.