Bring your own AI endpoint into a HubSpot workflow.
Save any OpenAI-compatible endpoint as an AI Connector, then run it from any workflow. Self-hosted models, OpenRouter, Azure OpenAI, Groq, or a gateway you already run. Your key is encrypted and never entered in HubSpot, and you only pay a flat 50 credits per successful run.
Works on the HubSpot tiers AI Studio supports, no Enterprise required
Classify the account tier on your own hosted model
Set Account tier to the outcome of Send to Custom AI
If it speaks the OpenAI chat schema, it works.
One connector calls POST {base URL}/chat/completions over HTTPS. The model id is free text, so a new model works the day it ships.
Self-hosted
vLLM, Ollama, LM Studio, TGI, or any server that speaks the OpenAI chat schema, behind a TLS tunnel.
OpenRouter
One key, hundreds of models. Route by vendor slug and pick the price or latency you want.
Azure OpenAI
Your Azure resource on the v1 endpoint, so inference stays in your tenant and your region.
Groq and friends
Groq, Together, Fireworks, DeepInfra, Mistral, and other hosted OpenAI-compatible APIs.
Gateways
Front a LiteLLM proxy, Portkey, or Cloudflare AI Gateway you already run for routing and spend control.
Your model of choice
Model id is free text, so new models work the day they ship. No waiting on a vendor list.
Four steps, and you test before you save.
The Playground and the Test connection button run the exact same request the workflow will. A green check means production will work. Both are unbilled.
01
Save a connector
In AI Studio, add an AI Connector with a name, base URL, API key, and default model. Your key is encrypted and never entered in HubSpot.
02
Test before you save
Run a live test against your endpoint from AI Studio. A green check means the same request will work in production. Tests are unbilled.
03
Drop the action in
Add Send to Custom AI to any HubSpot workflow. Pick the connector from a masked dropdown and write your prompt with CRM tokens.
04
Use the output
Save the response to a property or note, split it into fields, or branch on the status. Bill only 50 credits, only on success.
Sensible defaults out of the box: connector, model, prompt. Flip on Advanced setup when you want to tune the system prompt, temperature, reasoning effort, max tokens, or pass extra JSON parameters.
More than a blob of text back.
Every run returns structured outputs you can map, branch on, or inspect, so a failure routes cleanly instead of landing in a CRM property as error text.
AI response
The generated text, ready to save to a property, note, or the next step.
Status you can branch on
A clear status code (success, truncated, auth failed, timeout, model not found, and more) so a workflow can route on the result.
Error message
A plain, admin-actionable message when something goes wrong, instead of error prose written into a CRM property.
Split outputs
Split the response into up to five separate outputs to map into different fields.
Run details
Provider-reported tokens, finish reason, and duration, plus the full response JSON when you need it.
Timeouts you control
Pick 30, 60, 120, or 300 seconds per connector, so slow self-hosted models get the room they need.
Retry on the wobble
Transient connection errors and rate limits retry once automatically, honoring Retry-After.
Dead endpoints step aside
A connector whose last ten runs all failed is disabled automatically, so a broken endpoint stops burning workflow steps.
Billed at most once
If HubSpot delivers the same enrollment twice, the run executes and bills only once. Failed runs never bill.
Your model, your rules, inside HubSpot.
Choose where inference runs
Point the connector at Azure OpenAI in your tenant, or a model you host in the EU. Inference runs on the endpoint you chose, and your prompt reaches only that provider, never a third-party model service.
Run open models in regulated teams
Healthcare, finance, and legal teams can self-host Llama, Hermes, or Qwen and still trigger them from a HubSpot workflow.
Cut model spend
Route through OpenRouter or Groq to grab a cheaper or faster model for low-stakes steps, and pay the provider rate you picked.
Reuse your AI gateway
Already run LiteLLM or Portkey for routing, keys, and logging? Send to Custom AI plugs straight into it.
Fine-tuned and private models
Call a model you fine-tuned on your own data from any workflow, without exposing the endpoint in HubSpot fields.
Branch on the result
Every run returns a status output, so a workflow can retry, route, or notify on failure instead of writing error text into a property.
Your key stays out of HubSpot.
HubSpot workflow fields have no secret type, so anyone who can edit a workflow can read them. Your credentials live in an admin-only connector instead, and the workflow only ever sees a masked dropdown.
Keys never enter HubSpot
HubSpot workflow fields are visible to every workflow editor. Your API key lives in an admin-only connector instead, referenced by an opaque id.
Encrypted at rest
Secrets are stored with AES-256-GCM envelope encryption, write-only through the API, and decrypted only at execution time.
No third-party gateway
Your prompt goes straight to the endpoint you chose over TLS. AI Studio never routes it through a third-party model gateway.
Hardened egress
Every request is HTTPS only, pinned to the resolved IP, blocked from private and metadata ranges, with no redirects and a strict size cap.
Metadata-only logging
By default AI Studio logs only run metadata: connector, host, model, and status. Prompt and response logging stays off unless you turn it on.
You pay your provider
Inference runs on your account with your key. AI Studio charges a flat 50 credits per successful run, and failed runs cost nothing.
A custom base URL, on a HubSpot tier you already pay for.
HubSpot's own custom LLM action is Enterprise-only and limited to a short list of named vendors, with no way to point at a base URL of your own. Send to Custom AI takes any OpenAI-compatible endpoint you control, on the tiers AI Studio supports. Self-hosted for data residency, OpenRouter for cost, Azure for your tenant. Same workflow builder, same tokens, same outputs you already map.
Send to Custom AI questions.
Which endpoints work with Send to Custom AI?+
Where does my API key live?+
Does my CRM data leave my portal?+
What does it cost?+
Do I need HubSpot Enterprise?+
Can I use it in Breeze agents too?+
What happens when my endpoint is down or slow?+
Point a HubSpot workflow at your own model.
Save a connector, test it live, and run it from any workflow. Your key stays encrypted and out of HubSpot.
- Any OpenAI-compatible endpoint you control
- Keys encrypted, never entered in HubSpot
- Flat 50 credits per successful run, failures free
Start free trial