Ollama streaming chat built with just urllib, no pip installs

The post is aimed at a common friction point: Ollama ships an official Python library, but some scripts are better off without it. By importing only urllib and json, the author avoids pip entirely, which suits single-file tools or machines where adding packages is a hassle. The script targets the local endpoint at http://127.0.0.1:11434/api/chat with a small qwen2.5:0.5b model, so the whole loop stays visible in about 20 lines.

The mechanism rests on NDJSON: with stream set to true, each line is a self-contained JSON object holding a piece of message.content and a done flag. Iterating over the response and printing each piece with flush is what produces the typewriter effect in a terminal, though redirecting to a pipe collapses it into one block. The author notes that json.loads accepts the raw bytes directly, since Python 3.6 and later detect UTF-8 automatically, sidestepping decoding pitfalls.

The caveats are practical rather than conceptual. The first request waits on model loading, measured at roughly 30 seconds for the 0.5B model, which matters if you set your own timeout. And once streaming works, conversation history and token-stat formatting become your own responsibility, so the author suggests the official library when that overhead grows; the point here is running with zero dependencies.