Integrating Ollama into a native desktop app: field notes

GhostNote talks to Claude, GPT and Gemini through their cloud APIs, and to Ollama for people who want everything local. Wiring in the local provider was the easiest integration of the four — and the strangest, in ways worth writing down.

Published 21 August 2026 · GhostNote runs on Windows and macOS

The happy path is plain HTTP to localhost

Every cloud provider integration is HTTPS, keys, and an auth dance. Ollama’s entire integration is a POST to 127.0.0.1:11434. No key, no auth, no TLS — the server is on your own machine and speaks plain HTTP. After writing three cloud auth flows it feels almost wrong. It isn’t; it’s just what local means.

Sampling params live in a nested options object

Coming from OpenAI-style APIs you’ll reach for a top-level temperature field, and Ollama will silently ignore it. Sampling parameters belong inside an options object in the request body. GhostNote sets temperature: 0.2 there — short factual answers want the most likely completion, not creative sampling — and it’s an easy thing to miss if you’re pattern-matching from the cloud providers.

Streaming is NDJSON — and we turn it off

Ollama streams by default, one JSON object per line, rather than the SSE format the cloud APIs use. GhostNote deliberately sends stream: false and takes the whole response in one shot: our answers are short, and one parse of one JSON object beats maintaining a line-delimited streaming reader. If your responses are long, stream; if they aren’t, simplicity wins.

The failure mode is absence

Cloud providers fail loudly, with status codes and error bodies you can switch on. Ollama fails by simply not being there — the user hasn’t installed it, hasn’t started it, or hasn’t pulled the model, and your request gets nothing at all. The real integration work wasn’t parsing errors; it was turning an empty response into a message a person can act on: install Ollama from ollama.com, run ollama pull <model>, make sure it’s running.

Why bother

Because local is the whole point for a real set of users: no API key, no usage bill, and nothing leaving the machine. GhostNote’s screenshot-and-ask flow pointed at a local model works offline entirely — the same overlay, running against weights on your own disk. How to set that up end to end is covered in running AI completely offline and on the offline page.

See pricing