How to connect n8n to n4n.ai for multi-model workflows

Most n8n automations stall when a single task needs different LLMs for different substeps. To connect n8n to n4n.ai multi-model workflow, point n8n’s HTTP Request node at the OpenAI-compatible endpoint and pass the model name per request—no dedicated n8n integration needed. You get 240+ models behind one credential and a uniform request shape.

Step 1: Provision credentials and base URL

Generate an API key from your gateway account. The base URL is a single OpenAI-compatible endpoint; export it as BASE_URL in your shell or n8n environment variables. Store the key in n8n as a generic HTTP Header Auth credential named Authorization with value Bearer <key>.

Test connectivity before touching n8n:

curl $BASE_URL/v1/models \
  -H "Authorization: Bearer $N4N_API_KEY" | head -c 200

A JSON array of model ids confirms the credential works. In n8n, create the credential once and reuse it across every model node. This avoids scattering secrets across workflows.

Step 2: Scaffold the workflow with a manual trigger

Start a blank workflow. Add a Manual Trigger node. Immediately after, add an HTTP Request node named Call Model A.

Configure the node:

• Method: POST
• URL: {{ $env.BASE_URL }}/v1/chat/completions
• Authentication: select the header credential from Step 1
• Send Body: true
• Body Content Type: JSON
• Body (JSON tab):

{
  "model": "openai/gpt-4o-mini",
  "messages": [
    { "role": "user", "content": "Summarize: {{ $json.inputText }}" }
  ],
  "max_tokens": 256
}

The {{ $json.inputText }} expression pulls from whatever upstream node you later attach. Keep the payload minimal; the gateway forwards unknown fields but some models ignore them.

Step 3: Parameterize the model per node

A multi-model flow means different nodes call different models. Duplicate Call Model A, rename to Call Model B, and swap the model field:

{
  "model": "anthropic/claude-3.5-sonnet",
  "messages": [
    { "role": "user", "content": "Extract entities: {{ $json.inputText }}" }
  ],
  "max_tokens": 512
}

Because the endpoint speaks the OpenAI chat format, n8n treats both calls identically. You avoid branching auth logic or maintaining separate base URLs. If you later need a dynamic model, replace the literal with {{ $vars.modelName }} and set the variable at workflow runtime.

Step 4: Pass routing directives and cache hints

When you need to pin a provider or exploit prompt caching, add route and cache to the body. The gateway honors client routing directives and forwards provider cache-control hints, so you can force a specific backend or set TTL without custom headers:

{
  "model": "meta-llama/llama-3.1-70b-instruct",
  "messages": [{ "role": "user", "content": "{{ $json.inputText }}" }],
  "route": { "provider": "groq" },
  "cache": { "ttl": 300 }
}

n8n sends this verbatim. No extra node config. This is the cleanest way to control cost and latency from inside the workflow logic.

Step 5: Merge outputs and branch on results

Add a Merge node (mode: append) after both model calls. Then add an IF node to validate responses:

// IF node condition (n8n expression)
{{ $json.choices[0].message.content.length > 0 }}

If you want a third model as arbiter, chain another HTTP Request node with model: "google/gemini-1.5-pro" and feed both prior outputs into its messages array. Use a Code node to assemble the conversation:

const a = $items("Call Model A")[0].json.choices[0].message.content;
const b = $items("Call Model B")[0].json.choices[0].message.content;
return [{ json: { messages: [
  { role: "user", content: `A says: ${a}\nB says: ${b}\nReconcile.` }
]}}];

This pattern scales to any number of models without changing the transport.

Step 6: Handle rate limits and fallback

Providers throttle. n4n.ai automatically falls back when a provider is rate-limited or degraded, so you can skip custom retry loops for model unavailability. Still, configure n8n node retries for transport errors:

• Settings → Retry On Fail: true
• Max Tries: 2
• Wait Between Tries: 2000ms

If the gateway returns HTTP 200 with finish_reason: "error", parse it downstream and route to an alert. Do not assume non-200 is the only failure mode.

Step 7: Meter usage and track cost

The response includes a usage object. Capture it:

{
  "usage": {
    "prompt_tokens": 42,
    "completion_tokens": 17,
    "total_tokens": 59
  }
}

Map {{ $json.usage.total_tokens }} to a Google Sheet or Postgres node. In a high-volume workflow, write these rows asynchronously with a separate branch to avoid blocking the main path.

Step 8: Verify the workflow end-to-end

Trigger the manual execution with a sample inputText from a Set node. Expect two model responses merged. Check the execution log: each HTTP node should show 200, and the Merge node should output an array of two items.

Add a verification Code node after Merge:

const items = $input.all();
if (items.length !== 2) throw new Error("expected 2 model outputs");
for (const i of items) {
  if (!i.json.choices?.[0]?.message?.content) throw new Error("empty completion");
}
return items;

If it passes, you have a working connect n8n to n4n.ai multi-model workflow. From here, swap models via workflow variables, add a Switch node to pick models by input type, or parallelize calls with a Loop node.

Caveats

n8n expressions can break JSON if you concatenate unescaped strings. Use {{ JSON.stringify($json.text) }} inside message content when needed. The OpenAI-compatible surface does not support every provider-specific extension; test a new model id with curl $BASE_URL/v1/chat/completions before wiring it into n8n. Keep credentials in n8n’s vault, not in node bodies.