How to Connect a Model to Nod by n8n

Nod ships without a model — everything it does runs on one you connect in Settings → Models. You have three kinds of route: sign in with a ChatGPT Plus or Pro subscription (no API key needed), paste API keys from providers like OpenAI, Anthropic, or Google, or run a local model on your own machine. Nod itself is free; the model bill, if any, is between you and the provider. This page walks through each route, the two sign-in exceptions that trip people up, and what the official docs require of local models.

Which route should you pick?

You aren't locked in: Nod supports multiple connections at once, even two accounts on the same service, and you can assign different models to chat, schedules, and autonomous work.

Option 1: Sign in with ChatGPT Plus or Pro

This is the only subscription sign-in Nod offers, and for most people it's the fastest start.

  1. Open Settings → Models and add a new connection.
  2. Choose OpenAI / ChatGPT.
  3. Choose the sign-in option and log in with the ChatGPT account that holds your Plus or Pro plan.
  4. Confirm the model selection. Per the official docs, the initial choice for a new ChatGPT connection is GPT-6.2 Sol (the docs note these defaults can change between versions).

The two exceptions. Don't assume other subscriptions work the same way — per the official docs, they don't. A Gemini app subscription cannot be used to sign in to Nod, because Google doesn't allow third-party assistants to reuse it; Gemini needs an API key instead. And Claude must be connected by API key, because Anthropic's authentication conditions require it. ChatGPT is the one place where "I already pay for the app" carries over.

Option 2: Connect with an API key (nod api key)

  1. Create or copy an API key from your provider's console.
  2. In Nod, open Settings → Models, add a connection, and choose the provider.
  3. Paste the key, then select the model you want that connection to use. The docs list initial choices per provider — for example Claude Opus 5.5 for Anthropic and Gemini 4 Argon for Gemini — and note they may change between versions.

The provider list is long. Named first-class connections in the docs include OpenAI, Anthropic, Google Gemini (AI Studio key), OpenRouter, Google Vertex AI, Amazon Bedrock, Azure OpenAI, Cloudflare Workers AI / AI Gateway, and any OpenAI-compatible endpoint where you supply the server address. Also named: DeepSeek, Groq, Mistral, xAI, Together AI, Fireworks AI, Cerebras, Hugging Face, Moonshot AI, MiniMax, NVIDIA, and Z.AI — the product page counts 30+ providers in total.

Billing is direct: requests go to the service you connected, and the provider charges you, not n8n. Nod shows each task's cost as it goes and issues a receipt at the end.

Option 3: OpenRouter

OpenRouter fronts many providers' models behind one account. Per the docs, you can connect it with an API key or by signing in through the browser, and its initial-choice model is GPT-6.1 Sol. Pick this route if you like switching models often and don't want a separate key — and a separate bill — for every provider.

Option 4: Run a local model (no API key at all)

Nod can run on a model hosted on your own machine through Ollama, LM Studio, vLLM, or llama.cpp, connected as an OpenAI-compatible endpoint. No key, no per-task fees. But the official docs set a hard bar, and it's higher than most people expect:

Two caveats from the docs and product page. First, small local models are unreliable at Nod's safety checks — the check that decides whether an action needs your approval is only as good as the model making it. Second, when you use a local model, the trust check itself runs on a known small model or a decision service; that component can still generate a small API cost, and the input to the check is sent to that service. Local usage is recorded in Nod's Usage log as "No price" and doesn't count toward your monthly spending limit.

Using more than one model

One connection is enough to start, but it isn't the ceiling. Per the docs, you can keep several connections live at once — including two accounts on the same service, such as two ChatGPT accounts — and assign different models to different kinds of work: one model for chat, another for schedules, another for work Nod starts on its own, including a lighter, cheaper model on the same connection for simple jobs. A sensible setup is a capable model where judgement matters (anything involving the safety check or other people) and a light model for routine, read-only work.

What you'll actually pay

Nod is free and open source; the only money in play is model usage. Three shapes:

Whichever route you choose, spend a moment on model quality for your main connection. In n8n's own published testing of the safety check, its top models judged all 240 test decisions correctly while a small model let 8 unsafe actions through — the docs' advice is simply to pick a capable model.

Next steps