Connect Make.com to Self-Hosted AI on GPU

What You’ll Connect

After this guide, your Make.com scenarios will route AI requests through your own GPU server — no API costs, no rate limits. Every automation that needs text generation, summarisation, or classification calls your private vLLM or Ollama endpoint directly from Make.com’s visual workflow builder.

Make.com’s HTTP module sends requests to your dedicated GPU server the same way it calls any REST API. The visual canvas makes it straightforward to map trigger data into prompts, parse completions, and feed results into hundreds of connected apps — from CRMs to spreadsheets to messaging platforms.

HTTP Module (POST) –> GPU Server –> LLM Inference (webhook, email, /v1/chat/completions Nginx proxy dedicated GPU schedule, app) + auth hardware | | Downstream <-- Parse JSON <-- HTTP response <-- Model completion modules extract text 200 + JSON body returned -->

Prerequisites

• A GigaGPU GPU server running an LLM behind an OpenAI-compatible API (self-host guide)
• HTTPS with a valid certificate on your GPU endpoint (Nginx reverse proxy guide)
• A Make.com account — the free tier supports HTTP modules for testing
• An API key on your inference server for request authentication (security guide)

Integration Steps

Open Make.com and create a new scenario. Add your trigger module — this could be a webhook, a scheduled interval, a new email, or any app event that provides data for the AI to process.

Add an HTTP – Make a Request module after the trigger. Set the method to POST and the URL to your GPU server’s completions endpoint: https://your-gpu-server.gigagpu.com/v1/chat/completions. Under Headers, add Authorization: Bearer YOUR_KEY and Content-Type: application/json. Set the body type to Raw and content type to JSON.

Build the JSON body using Make.com’s data mapping to inject trigger values into the prompt. After the HTTP module, add a JSON – Parse JSON module to extract choices[0].message.content from the response. Feed that parsed text into any downstream module — a Google Sheets row, a Slack message, or an email reply.

Code Example

Configure the HTTP module’s raw JSON body. Use Make.com’s curly-brace syntax to map dynamic data from the trigger into the OpenAI-compatible format:

{
  "model": "meta-llama/Llama-3-70b-chat-hf",
  "messages": [
    {
      "role": "system",
      "content": "You are an automation assistant. Be concise and structured."
    },
    {
      "role": "user",
      "content": "Classify this support ticket and suggest a response: {{1.ticket_body}}"
    }
  ],
  "max_tokens": 400,
  "temperature": 0.2
}

// HTTP Module Settings:
// URL: https://your-gpu-server.gigagpu.com/v1/chat/completions
// Method: POST
// Headers: Authorization = Bearer sk-your-key
// Body: Raw JSON (above)
// Parse response: Yes

Testing Your Integration

Run the scenario once using Make.com’s “Run once” button. The HTTP module should return a green checkmark with the full JSON response. Click into the module output to verify choices[0].message.content contains a valid completion. The JSON Parse module should cleanly extract the text field.

Test the full pipeline by triggering the scenario with real data. Verify that downstream modules receive the AI-generated text and process it correctly. Check execution logs for any timeout errors — Make.com’s default HTTP timeout is 40 seconds, which should accommodate most inference requests.

Production Tips

Make.com charges by operation count, not by data volume, so routing AI through your own GPU eliminates the double cost of paying both Make.com and a third-party AI provider. Every scenario that previously used an OpenAI module can switch to an HTTP module pointed at your self-hosted API with identical payloads.

For error handling, add a Router after the HTTP module with an error-handling path. If the GPU server returns a 500 or times out, queue the request for retry rather than losing the trigger data. Make.com’s built-in retry and error handling works well for transient failures.

To scale throughput, use Make.com’s parallel execution feature alongside a vLLM deployment configured for concurrent request batching. This lets multiple scenario runs hit your GPU simultaneously without queuing. Explore open-source LLM hosting options, browse more tutorials, or get started with GigaGPU dedicated GPU hosting to power your Make.com automations.