Different tasks, different models

[gadgetb0y]

Can different tasks use a specific model? For instance, I would want to use Claude Sonnet 4.6 for interactive tasks but I would prefer background tasks (morning briefings, heartbeat checks) to use my local LLM, where slower processes are fine. OpenClaw set this up for me automatically just by chatting with the bot.

[freddy-schuetz]

Hey @gadgetb0y, great question! This is actually already possible today. Just not automatically, but with a few clicks in the n8n UI.

How to do it:

1. Create a second credential in n8n for your local LLM (e.g. Ollama)
2. Open the background workflows you want to run on the local model (e.g. Heartbeat, Memory Consolidation, Background Checker)
3. Swap the LLM sub-node inside the AI Agent node to your local model
4. Done: interactive chat uses Claude, background tasks use your local LLM

Since each workflow has its own independent AI Agent node, you can mix and match models freely per workflow.

Why not build automatic routing?

n8n already gives you per-workflow model control natively. Building a DB-based routing system on top would add complexity without real benefit. The UI approach is straightforward and takes about 5 minutes.

The setup.sh script does set all workflows to the same provider during initial deploy, but after that you are free to customize individual workflows as needed.

[gadgetb0y]

Thanks, @freddy-schuetz.

I'm not deep into n8n just yet and didn't think about setting up workflows to use a specific LLM - I was using the OpenClaw mental model. This will work well.

[dankerthrone]

I'm very new to all this as well. On a productive system where I only have n8n installed (not n8n-claw), I cobbled together this workflow:

you can see two LLM nodes, both of which currently address mistralai/mistral-small-2603 via my openrouter API. The whole workflow is basically "get newly scanned and OCR'ed pdf doc from paperless-ngx, send it to Mistral to semantically improve OCR text (such as "l0ng"-->"long"), feed it back into paperless-ngx, take the entire text you just excerpted and send a summary and actions items if applicable to me via Telegram". I don't want frontier models to do that because it's too expensive. This is something I can cheap out on, so I delegate these tasks to Mistral. Now to figure out how to incorporate this workflow into my n8n-claw instance and how to write the scanned text into n8n-claw's memory ... 🤔

[gadgetb0y]

@dankerthrone Local, self-hosted LLMs have changed the game for me. I don't worry about privacy since the data never leaves my LAN and I don't worry about token costs.

No, it's not as fast as a frontier model running hot on an H200 in a data center, but for the type of task that you describe, local LLMs are perfect because you're not sitting there waiting for a response - it runs as a background task.

I picked up an M1 Ultra Mac Studio with 128 GB of Unified Memory on eBay and it does a great job. You can find them on eBay for ~$2K. I've easily saved that much in token costs since I purchased it last July.