The software expected a home.
Ollama was not shaped around this exact portable workflow. Packaging it meant finding which paths and machine-level environment assumptions mattered.
Evidence file 03 / local AI package
Portable Ollama
A zip-based local LLM setup that can run from a folder or USB drive without asking the host machine for a whole installation ritual.
What it was actually solving.
Portable Ollama was an attempt to make local LLMs less fussy to share: unzip the folder, run the launcher, and keep the host system as clean as possible.
Move the assumptions.
The launcher controls the runtime environment so model storage and supporting paths resolve around the package folder instead of a conventional machine-wide setup.
Observe the black box.
Changing software you do not own starts with tracing its assumptions. Control the environment, test the seams, and package the result so the next person does not have to repeat the investigation.
Illustrative route map / not live telemetry
Folder to model.
The useful path is deliberately short: unpack, launch, resolve local paths, and bring the model server up without a machine-wide detour.
More from the bench.
Utilities built to remove setup or put a useful control surface in front of it.