The Newsletter Research Desk has one step that calls an LLM: it summarizes your ranked sources. That endpoint is a placeholder by default, and people assume it means a paid cloud API. It doesn't have to. If you run a model locally with Ollama, you can keep the entire workflow — and your data — on your own machine.
The setup, in plain terms
- Run Ollama locally and pull a model you like (a small instruct model is plenty for summaries).
- Ollama exposes an HTTP endpoint on your own network.
- In the workflow's 'Summarize with your LLM' node, point the URL at that endpoint and shape the JSON body to match Ollama's API.
- No external API key, no data leaving your network.
If the model isn't reachable, the workflow doesn't fall over. The summarize node is set to continue on failure, and the next step has a deterministic fallback that still produces an approvable outline. You lose the richer summaries, not the run.
Why bother going local
- Privacy: source notes and drafts never leave your machine.
- Cost: no per-token bill for routine summarization.
- Control: you pick the model and version, and nothing changes under you.
Self-hosting knowledge shouldn't have a language barrier — or a vendor lock-in.
The trade-off is quality and speed versus a frontier cloud model. For ranking and summarizing your own collected sources, a local model is usually more than enough — and the approval gate means a human reads the output anyway.