review-relay
v0.15.0Docs

Your agents judge every pull request.The lowest score counts.

Seat Claude Code, Codex, Gemini CLI, Copilot, or any of 13 supported agent CLIs, as many as you like. review-relay starts them on your machine the moment a review begins on GitHub, then posts one comment with their findings and a 1–5 merge confidence score.

5
4
5

acme/billing #482

Retry failed invoice webhooks

4/5

Merge confidence

Lowest of 3 judges on 9f3c2a1

Example panel on a fictional repo with three configured agents.
$ curl -fsSL https://github.com/Tyru5/review-relay/releases/latest/download/install.sh | bash

Installs a standalone binary and verifies its checksum. Run it again to update.

The scoresheet

Every judge scores the same six dimensions from 1 to 5. Findings cap the score, so the number never outruns the problems. The comment’s headline is the lowest judge.

Set the severity of the retry-loop finding

Headline score 4

A minor finding leaves the score alone.

Retry failed invoice webhooks

acme/billing #482 · 9f3c2a1

Example review on a fictional repo with three configured agents.

DimensionAgent AAgent BAgent C
Correctness545
Security555
Code quality445
Standards544
Blast radius444
Testing534
Merge confidence545

Findings (2)

  • minorsrc/invoices/retry.ts:57Retry loop has no cap on total delay (Agent A, Agent B)
  • minorsrc/invoices/retry.test.tsNo test for the exhausted-retries path (Agent B)

Order of events

Nothing to click. The panel convenes on its own, and GitHub events reach your machine through gh webhook forward, so there is no public URL to host.

  1. A review starts

    Greptile or CodeRabbit starts reviewing, a pull request is opened or marked ready, or a collaborator comments @review-relay. Each repo picks its trigger: auto, greptile, coderabbit, or github.

  2. The commit is checked out locally

    review-relay fetches the PR head into a temporary worktree of your own clone, so the panel reads real code using your own agent logins and keys.

  3. The panel reviews in parallel

    Every agent on your panel reviews against the same rubric: your repo's conventions, lint config, and the callers of changed code. Each CLI runs headless and read-only, locked down by its own flags rather than by the prompt, and config files a PR could use to run code through it are removed from the worktree first.

  4. One scored comment lands

    Findings are merged across judges and linked to exact lines. Later reviews edit the same comment. Each commit is reviewed once, and drafts wait until they are ready. Turn on more per repo: a review-relay commit status branch protection can require, new findings on their lines with one-click suggestions, and a ping on your desktop, ntfy, or Slack when a score comes in low.

Score it before the PR.

No pull request yet? review-relay review puts the branch you have checked out in front of the same panel, against the same rubric. Wait for the score in your terminal, or send it to the background and watch it in the TUI. It needs no PR, no GitHub access, and no daemon, and posts nothing.

$ review-relay review --base origin/main --min-score 4$ review-relay review -d
  • Same rangeThe commits your branch adds over origin/main, or any base you pass, as the pull request would show them.
  • Same scoreThe panel, lockdown, caps, and lowest-score headline of a PR review.
  • Keep workingreview -d returns at once. The review shows up in review-relay tui next to your PR reviews.
  • Gate on it--min-score 4 exits 1 under a 4. hook install runs it on every git push. --json for tools.
  • From your agentreview-relay mcp serves the same review over MCP, so Claude Code, Codex, or any client can score the branch it is working on and fix the findings before it opens the PR.

Before the panel sits

review-relay drives tools you already use. Have these signed in first.

  • gh signed in with access to the repos you watch, plus its webhook extension:
    $ gh extension install cli/gh-webhook
  • At least one of these agent CLIs, signed in or set up with an API key. Claude Code and Codex are the defaults.
    • Claude Codeclaude
    • Codex CLIcodex
    • Augment Auggieauggie
    • GitHub Copilot CLIcopilot
    • Factory Droiddroid
    • Gemini CLIgemini
    • Grok Buildgrok
    • Hermes Agenthermes
    • Kilo Code CLIkilo
    • opencodeopencode
    • pipi
    • Qwen Codeqwen
    • Mistral Vibevibe
  • A local clone of each repo you want reviewed

setup finds the GitHub clones under your home folder so you can pick the repos to watch, then lists the agent CLIs on your PATH so you can pick your panel and each judge’s model and effort. It writes ~/.review-relay/config.json, start keeps watching, and tui shows the jobs as they run, with each PR’s state on GitHub and what every review cost:

$ review-relay setup$ review-relay start -d$ review-relay tui

Reviews run from your machine, and your code goes to whichever providers your chosen agents and models use.

Set the rules.

The docs cover your first review and every setting after it.

Choose models and effort, route different pull requests to different panels, decide when the relay runs, gate merges on the score, and connect your agents over MCP. Find the defaults, working config examples, and help when a review does not arrive.

Seat your panel.

Free to use. One command installs a standalone binary for macOS, Linux, or Windows.

$ curl -fsSL https://github.com/Tyru5/review-relay/releases/latest/download/install.sh | bash