When a team says it wants a self-hosted AI code reviewer, the real question is where the model runs, not which repo to clone. The tool is the easy part. The model placement is the decision, and it decides how much you actually gained.
PR-Agent's current repository is a clean place to see the distinction. It is now owned by the PR-Agent org on GitHub, donated by Qodo, and the README is explicit that this open-source project is not the Qodo commercial offering. It says it can run on GitHub, GitLab, Bitbucket, Azure DevOps and Gitea, and deploy as a GitHub Action, Docker, or a CLI. On the model side, it routes any model reachable through LiteLLM, which includes local runners like Ollama alongside hosted OpenAI, Anthropic, Google, DeepSeek and Mistral endpoints. (PR-Agent repository, checked 2026-09-14)
So the honest framing is three options, and they are not the same thing.
Hosted model, self-hosted app. The reviewer code runs on your infrastructure, but the LLM calls go out over an API. You control the workflow, the prompt config and the git platform integration. PR-Agent's README states this plainly: if you self-host with your own OpenAI key, it is between you and OpenAI on data. You keep the pipeline, not the model.
Local model, local app. Point the same self-hosted tool at a model running on your own hardware through Ollama or a LiteLLM-routable local endpoint. Code never leaves the machine. This is the air-gapped setup, and it is the only one where the word "private" is doing real work. The trade is visible: you are now managing the hardware and the model, and your review quality tracks the local model you are willing to run.
Hosted everything. The tool's vendor runs both the app and the model. This is what "free" cloud tiers and most commercial review tools are. It is not self-hosted at all, which matters if the point was to keep code off third-party infrastructure.
The recurring mistake in the "best self-hosted tool" search is treating these as one product category. They are three different operations with three different failure modes. A team that wants its next git push to run a reviewer on internal infrastructure but is fine calling a hosted LLM has no reason to change its model choice, only its tooling layer. A team chasing a data boundary has to also answer what model it can run locally that produces reviews it trusts. The tool cannot fix the model.
None of the third-party write-ups I checked give you a verifiable way to rank these against each other, and their suggestions are mostly prompts rather than evidence. The place to look first is the tool's own repository and docs, because the model-placement question is answered there in source, not in a listicle. (PR-Agent deployment and models, checked 2026-09-14)
Top comments (0)