Self-hosting an AI code reviewer hinges on where the model runs, not the tool

Teams evaluating self-hosted AI code review tools face a critical but often overlooked choice: whether the underlying language model runs on their own hardware or on a third-party API. PR-Agent, an open-source reviewer now maintained by its own GitHub org after being donated by Qodo, illustrates three distinct deployment configurations — hosted model with self-hosted app, fully local model and app, or fully hosted by a vendor. Only the fully local setup, using tools like Ollama routed through LiteLLM, ensures code never leaves internal infrastructure and gives the term 'private' genuine meaning. Teams comfortable calling an external LLM API need only reconsider their tooling, while those with strict data boundaries must also evaluate which local model delivers trustworthy review quality. The answer to the model-placement question is best found in a tool's own repository and documentation, not in third-party comparison articles.