What ReviewStage is

ReviewStage is a self-hosted web dashboard that drafts pull-request reviews for you and lets you post them, one click at a time, under your own GitHub name. It is built on Claude Code, runs on your own Claude subscription, and never posts anything you did not select.

It is a reviewer’s assistant, not a review bot. The distinction is the whole product.

The one-minute model

a PR needs your review
  └─ you open it in ReviewStage (or a Slack card brings you there)
       └─ click Start review: effort, optional focus note, model
            └─ the agent checks out the branch in a worktree and reads the real diff
                 └─ review.json — findings with file:line, severity, body   (a few minutes for a typical PR)
                      └─ you tick the findings worth posting, edit any wording
                           └─ Post to GitHub — a plain COMMENT review, as you
                                └─ Approve — a separate click, also as you

Three properties hold at every step:

1. The review step cannot write to GitHub. The script that runs the agent has no write path at all; it produces a file the dashboard renders.
2. Every write is a signed-in human’s click, with that human’s own token. There is no bot account. If a comment carries your name, you chose it.
3. Posting defaults to a COMMENT review. The agent never requests changes; a reviewer can tick Request changes on the post form. Approval is a second, deliberate click.

What you get

• A queue of PRs awaiting your review, with tabs for To review, Reviewed, Posted and Approved, so finished work leaves the working set.
• A PR page with the agent’s assessment, a plain-language explainer of what the PR does, its analysis, and the findings as editable cards.
• A learnings loop: what you drop or reword is remembered and fed into the next review of that repository.
• Skills: run reviews with a shared team skill or your own, score each by how often its findings are kept, and version the team standard.
• QA guides: a tester-ready P0/P1/P2 test plan built from the same diff and review threads.
• Insights: reviews, tokens, keep rate, agreement across reviewers, cycle time.

What it deliberately does not do

• It never posts automatically, on a schedule, or on push.
• It never posts as a bot or a shared identity.
• It never requests changes.
• It does not run on every PR. A review is a full agent run and costs real tokens; you choose which PRs earn one.

Next

• Install — Docker Compose in a few minutes, or from source on a Linux server.
• Your first review — from a PR URL to a posted comment.
• Team mode — one server, every reviewer, their own name.