AI code review for GitHub PRs

AI code review, anchored to real lines

A provider-agnostic GitHub Action — Claude or OpenAI — that posts inline review comments with one-click suggestions, and runs CVE and secret scanning in parallel, all in a single PR review.

Get started →

View on GitHub

Everything in one review

An AI reviewer plus deterministic scanners, sharing the same severity pipeline and posting once.

Anchored inline comments

Suggestions land on the exact diff line — not a wall of text at the bottom of the PR.

No hallucinated lines

The post tool validates every (file, line) against the real diff before accepting, and hints the agent toward valid lines on rejection.

Dependency CVE scanning

Parses changed lockfiles and queries OSV.dev for known CVEs — npm, yarn, pnpm, and pinned pip requirements.

Secret detection

Scans added lines for ~14 high-confidence credential patterns — AWS keys, GitHub PATs, Slack, Stripe, PEM keys — masked before posting.

Claude or OpenAI

Defaults to Claude Sonnet. Swap to Haiku for cost, Opus for depth, or a GPT / o-series model — provider is inferred from the id.

Noise control

Severity floor plus per-file and per-PR comment caps keep reviews signal-dense instead of overwhelming.

Repo-aware context

Reads .vor.yml, CLAUDE.md, AGENTS.md, and custom docs so findings respect your conventions.

Sticky reviews

Re-run any time. Prior agent reviews are dismissed so you only ever see the latest pass against HEAD.

Works with your stack

The AI review reads any language — no per-language setup. Deterministic scanners add zero-token depth wherever your toolchain runs.

Any language

The agent reasons over any diff the model understands — no language list to maintain.

JavaScript & TypeScript

Your repo’s ESLint, tsc —noEmit, and knip, surfaced inline.

Python

Ruff, run against the .py files your PR changes.

Dart

dart analyze on changed .dart files.

GitHub Actions

actionlint flags workflow and shell mistakes in your CI.

Custom rules

Bring your own Semgrep ruleset under .vor/semgrep-rules (opt-in).

Dependency CVEs

npm (package-lock, yarn, pnpm) and PyPI ( requirements.txt) via OSV.dev.

Secrets

~14 credential patterns — any language, any file.

Linters run only when the tool is available in your repo, so they stay silent on stacks they don’t apply to.

60-second setup

Drop one workflow file in your repo, add your API key as a secret, and trigger a review by hand.

.github/workflows/vor.yml

name: Vor
on:
  workflow_dispatch:
    inputs:
      pr_number:
        description: 'PR number to review'
        required: true

permissions:
  contents: read
  pull-requests: write

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - uses: driches/vor@v0
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          pr_number: ${{ inputs.pr_number }}

Prefer OpenAI? Swap in openai_api_key and set model: gpt-4.1 — the provider is inferred from the model id. Then: Actions → Vor → Run workflow → enter the PR number. A sticky review appears within minutes.

What a review looks like

Findings land on the exact line, tagged by severity, with concrete suggestions and provenance.

src/handlers/upload.ts

- const dest = path.join(UPLOAD_DIR, req.body.filename) + const dest = path.join(UPLOAD_DIR, sanitize(req.body.filename)) await fs.writeFile(dest, buf)

V

IMPORTANT · security

Path traversal — req.body.filename flows into path.join unsanitized, so ../../etc/passwd escapes the upload dir.

Why it matters: an attacker controls the write path and can overwrite arbitrary files.

Suggested change

const dest = path.join(UPLOAD_DIR, sanitize(req.body.filename))

V

CRITICAL · security

lodash@4.17.20 is affected by prototype pollution.

via OSV · GHSA-jf85-cpcp-j695 — fixed in 4.17.21

Why this works when “ask the AI to review the PR” doesn’t

Output is actionable, not prose

The agent has no free-text output channel. Findings can only surface through the validated post_inline_comment tool.

Comments land inline

Reviews post via pulls.createReview with a structured comments[] array — path, line, side, and suggestion blocks.

No hallucinated lines

The validator rejects any line outside the real diff and returns the valid ranges as a hint, so the agent self-corrects.

Add it to your repo

Free, MIT-licensed, and runs on your own API key. Manual-trigger by default — no surprise token spend.

Get it on the Marketplace

Read the docs