A private alternative to Zapier's AI features
This page shows how to keep the data in an automation workflow's AI step away from third-party model providers. If your platform's AI step accepts a custom OpenAI-compatible endpoint, you point it at a model on your own DGX Spark. If it does not, you move that workflow to n8n, whose AI nodes do. Either way, the customer record, internal email or contract clause in the workflow goes to a machine you rent instead of to the platform's model provider.
You need:
- A running GPUwerk instance with an OpenAI-compatible endpoint: a console model preset with an API key that has the
inferencescope, or your own engine on a base image. - Admin access to the automation platform's AI step settings, or an n8n instance if you migrate.
- No SSH key. GPUwerk configures the machine, and the console's browser terminal works without one. Add your own key only if you want your own SSH client, scp or port forwarding.
What's actually happening in that AI step
An AI action in a hosted automation platform is an API call the platform makes for you to a model provider it works with. You write a prompt, the platform fills in your workflow's data, sends it, and passes the completion to the next step. That is a sensible default for a general product. It is a different matter when the workflow pulls CRM rows, personal data from a form or the text of a contract, none of which left the platform before the AI step was added.
You do not have to stop using automation platforms. For each workflow with an AI step, find out where that step sends its data, and whether you can change the destination.
Check whether your platform supports a custom endpoint
Support varies between platforms and changes between releases, so check the settings of the specific AI action. Look for:
- An "AI provider" or "model" dropdown that includes a custom or "OpenAI-compatible" option alongside the platform's built-in models.
- A field for a base URL or endpoint URL on that AI action, separate from the platform's own API key field.
- Whether that option is available on your current plan tier; a custom-endpoint setting is sometimes reserved for higher plans even when the underlying feature exists.
If your platform has this:
- Set the base URL to your Spark's endpoint plus
/v1. For a console preset, that is the HTTPS endpoint from the instance page (it looks likehttps://app-<instance-id>.gpuwerk.com/v1); for your own vLLM server,http://<name>.gpuwerk.com:8000/v1. Or use a LiteLLM gateway in front for per-workflow keys. - Put your inference-scope API key (or the gateway's virtual key) in the API key field.
- Set the model to the ID your endpoint reports at
/v1/models, for exampleqwen3-coder-30b. - Run the step once with test data. The output should come back from your Spark, and the rest of the workflow stays the same.
n8n as the self-hostable alternative
If the platform has no custom endpoint for its AI steps, move the workflow. n8n is the practical choice: a self-hostable trigger-and-action workflow builder whose AI nodes accept a custom OpenAI-compatible base URL. The n8n guide covers the setup end to end, from the OpenAI credential to which nodes to use for a completion or a tool-using agent.
n8n is not a drop-in replacement. Its set of app connectors is large but different, so an integration you rely on may not exist there. Choose it when keeping the AI step's data off third-party models matters more than matching every integration.
Migrating a specific workflow
Workflows with a trigger, one or more AI steps and one or more output actions move cleanly:
- Rebuild the trigger (webhook, form submission, scheduled poll) as a Webhook or Schedule Trigger node.
- Rebuild each AI step as a Basic LLM Chain or AI Agent node pointed at your Spark.
- Rebuild each output action (write to a sheet, post a message, update a record) with the matching n8n node. An HTTP Request node covers most services without a dedicated n8n integration.
- Run the same input through the old AI step and the new one and compare the output. A prompt tuned for the platform's model may behave differently on the model on your Spark. Switch over once the results match what the downstream steps expect.
What actually breaks
- No base URL field in the AI step. Cause: the action is hardcoded to the platform's own model. Fix: move the workflow to n8n.
- The custom-endpoint option is greyed out. Cause: it is reserved for a higher plan. Fix: check the platform's current pricing page before you plan around it.
- A downstream step breaks after the switch. Cause: the new model formats lists or adds preambles differently. Fix: adjust the prompt and re-test every step that parses the AI output.
- The migrated workflow lost an integration. Cause: n8n has no connector for that app. Fix: use an HTTP Request node, or keep workflows whose AI step handles no sensitive data on the original platform.