Roboviewer

Automated code review for merge requests, running entirely on your machine.

Point it at two branches and it comes back with a ranked list of problems worth fixing — not a wall of style nitpicks.

roboviewer develop

▸ feature/discount → develop: 12 files, 8 checklist items
• Correctness and logic errors: 3 findings (ok) · 41200 tokens · 62s
• Error handling: 1 findings (ok) · 38900 tokens · 55s
...
▸ Confirmed 4 of 11

  F001  [Blocker] src/cart.py:42 — race when the cart is updated concurrently
  F002  [Major] src/api.py:118 — the change breaks older clients

Confirmed 4 of 11 · 167100 tokens · 67% from cache
Report: .roboviewer/runs/20260730-172900/report.md

Findings come back in English. --language ru asks the model for another one without touching the prompts — see Output language.

The problem

Review arrives late, and tired. A merge request waits hours or days, and by the time someone opens it they are on their fourth review of the afternoon. The blocker ships under three comments about whitespace.

Hosted reviewers want your code. CodeRabbit, Copilot and the rest do good work, but every one of them means uploading the repository to somebody else's infrastructure. Inside a corporate network that is where the conversation ends.

They also want your forge. They plug into GitHub or GitLab and review what is already a merge request. Looking over your own branch before you open it, or reviewing a mirror that lives nowhere but your laptop, is not something they do.

Pasting a diff into a chat window invents problems. Given a few lines of context, a model will confidently report a missing nil check that sits twenty lines above the hunk. Nothing is ranked, nothing is verified, and a real blocker arrives in the same flat list as a naming preference.

What you get

Your code stays where it is. Roboviewer talks to any OpenAI-compatible endpoint — including a corporate gateway — so reviews never leave the network you already trust. Nothing is uploaded anywhere else, and there is no service to sign up for.

Any git repository, no integration. It reads two branches through plain git. No app to install on your organisation, no webhooks, no permissions to request.

Review before you open the MR. Run it on your own branch, fix what it finds, and let the humans spend their attention on design instead of on the bug you would have caught yourself.

A ranked list you can act on. Findings carry a severity, a file and a line, and a final judge pass throws out the ones that do not survive a second look.

Requirements

• Python 3.11+
• git
• An OpenAI-compatible endpoint with tool calling — the agents drive the review through tools, so a completions-only gateway will not work. roboviewer --check-provider tells you which side of that line yours is on.

Install

git clone git@github.com:axazeano/Roboviewer.git && cd Roboviewer
python3 -m venv .venv && .venv/bin/pip install -e .
ln -sf "$PWD/.venv/bin/roboviewer" ~/.local/bin/roboviewer

Configure

mkdir -p ~/.config/roboviewer
cp config.example.toml ~/.config/roboviewer/config.toml
export ROBOVIEWER_API_KEY=...
roboviewer --check-provider

Set provider.base_url and reviewer.model. Everything else has working defaults and is documented inline in config.example.toml. --check-provider makes a handful of targeted requests and names what is wrong — wrong auth scheme, a base_url missing /v1, a gateway that cannot do tool calling — instead of leaving you to infer it from eight agents failing at once.

The sections, the rule that --config replaces rather than layers, and what to do about rate limits: Configuration.

Use

roboviewer <target> [source]

The target branch is required. The source defaults to your current branch, and naming it explicitly lets you review someone else's branch without checking it out.

roboviewer develop                    # current branch into develop
roboviewer develop feature/login      # someone else's branch
roboviewer -C ~/projects/app develop  # a repository living elsewhere

Reports land in .roboviewer/runs/<timestamp>/, and --diff-only shows what would be reviewed without spending tokens. Every flag: Command line.

Documentation

docs/ also carries the tooling baseline and how the measurement corpus is built.

What it doesn't do

• It does not post comments on your merge request, and does not talk to GitHub or GitLab at all. Output is files on disk; in CI it is the pipeline that publishes them, from formats the forge already understands.
• It does not modify your code. The agents get read-only tools — read_file, grep, list_files, git_show — and nothing else.
• It does not replace a human reviewer. It catches the class of problem that survives a tired read; it has no idea whether the feature was worth building.

License

MIT — see LICENSE.