Source code
- Hosted
- Never
- Local
- Yes, on the runner
Plug and Play Reviewer
Built for solo devs shipping 2 to 3 pull requests a day
Private AI code review that stays on your machine
A hosted control plane takes GitHub events. Your local runner reads the diff and calls your model. The patch never leaves the machine you trust.
That boundary is not a promise on a slide. CI runs a schema check that fails if source, diffs, or provider keys could land on the hosted plane. See what we store in Privacy.
An open source PR reviewer built for private AI code review. It is AI code review self hosted on a laptop or a server you run.
See how a review movesHow a review moves
GitHub PR to hosted job to local runner to retrieval to review comment. The hosted box never holds the patch.
GitHub PR
The GitHub App receives the pull request event and the head SHA.
hosted job
The control plane stores job metadata only. It does not receive the diff.
local runner
Your runner claims the job, fetches the patch, and keeps the source local.
retrieval
Repo chunks join the packed diff before the model call.
review comment
A comment does not post until a human approves it.
Hosted vs local
Source, diffs, and your provider API key stay on the runner. Finding text may sit on the dashboard so you can read it before you approve a post.
Diffs
- Hosted
- Never
- Local
- Yes, packed for the model
Your LLM provider API key
- Hosted
- Never
- Local
- Yes, in the local key store
GitHub event metadata
- Hosted
- Yes
- Local
- Yes, to claim the job
Finding title and rationale
- Hosted
- Yes, so the dashboard can show it
- Local
- Produced here
What already works
Grounded findings, suggestion blocks, and the human gate are product rules, not slogans.
Local runner
The model call happens on your laptop or server. Source never goes to the hosted site. The hosted plane only ever learns that a job ran, which repository it was for, and what it cost.
Retrieval
The reviewer gets repo chunks with the packed diff. It is not a diff-only guess.
Grounding
A finding that cannot point at the packed hunk is dropped before anyone sees it.
Suggested fixes
An optional suggested_fix can post as a GitHub suggestion when the replacement applies.
Opt-in specialists
Extra reviewers run only when a repo turns them on. They are off by default.
Human gate
The model cannot mark a finding public. Posting stays a human action.
What teams can change
You can add agents and edit prompts per repository. Those settings stay private to that repo.
Repo A never receives Repo B prompts. This is permissioned team power, not a public dump.
Access shape
What stays free and what is for teams
A single person running a local reviewer can stay on a simple path. The free tier allows one GitHub user and one repository per installation.
Team controls, shared prompts, and extra repositories are the paid path. There are no prices on this page.
What the evals show
Every finding is scored by a second model call and anything ungrounded is dropped before you see it. The scorecard shows the measured result on a human-judged holdout, with the model and the sample it was run on. That holdout is seven human-judged cases from the Zod repository, so it is evidence, not a published baseline.
Open the scorecardHow to set it up
GitHub App
Install the GitHub App and pick the repositories it may read.
Hosted URL
Point the runner at the hosted URL. plug-and-play-reviewer.vercel.app is the live control plane.
Local runner
Install the runner on the machine that may see your source.
Your LLM provider API key
The key from OpenAI, Anthropic, Groq or whoever you use. It stays on the runner and never reaches the hosted database.
One-click Render still needs your own secrets and database. Railway starts from a repo import until a public template id exists.
Read the install docs