open-code-review vs midscene
Two of the top projects in Code review and testing, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
open-code-review
Go CLI that reviews Git diffs with an LLM agent and line-level comments
midscene
Vision-based E2E testing in natural language for web, mobile and desktop
Verdict
open-code-review ranks higher, 91 to 87. Of the 4 score parts we measure for both, open-code-review leads on 1, midscene on 1 and 2 are even.
- Adoptionopen-code-review leads by 20 points, 94 to 74: a wide gap.
- FreshnessEven at 100.
- Maintenancemidscene leads by 5 points, 93 to 88.
- Agent-readyEven at 70.
Pick open-code-review if you are
Teams adding LLM pull request review to local workflows or CI
- Deterministic file selection, bundling and rule matching; one sub-agent per bundle for large changesets
- Delegation mode lets the host coding agent use its own LLM, so no OCR API key
Pick midscene if you are
QA and frontend engineers writing E2E tests for web, mobile or desktop apps
- Locates elements from screenshots, so canvas, icon-only buttons and cross-origin iframes need no selectors
- Same agent API across web, Android, iOS, HarmonyOS and desktop
| What we compare | open-code-review | midscene |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 94, widely used | 74, popular |
| Momentum | Not scored for this kind | Not scored for this kind |
| Freshness | 100, active | 100, active |
| Maintenance | 88, healthy | 93, healthy |
| Agent-ready | 70, partly | 70, partly |
| Facts from GitHub and the README | ||
| Stars | 49.8k | 15.6k |
| License | Apache-2.0 (permissive) | MIT (permissive) |
| Last commit | Nov 2026 | Nov 2026 |
| Last release | Nov 2026 | Nov 2026 |
| Language | Go | TypeScript |
| Docker | No | No |
| GPU | Not needed | Not stated |
| arm64 or Apple Silicon | Mentioned | Mentioned |
open-code-review
Open Code Review (`ocr`) reads a Git diff, bundles related files into sub-agent review units, and sends them to a configurable LLM that can read files and search the repo. It outputs structured comments with line positions, and `ocr scan` reviews whole files without a diff. It runs locally or in CI (GitHub Actions, GitLab CI, GitFlic CI, Gerrit), with plugins for Claude Code, Codex, Cursor, Kimi Code and OpenCode.
Who it is for: Teams adding LLM pull request review to local workflows or CI
Strengths
- Deterministic file selection, bundling and rule matching; one sub-agent per bundle for large changesets
- Delegation mode lets the host coding agent use its own LLM, so no OCR API key
- Separate comment-positioning and reflection modules to improve line accuracy
- Resumable review and scan sessions, with a browser session viewer
Weaknesses
- Requires Git 2.41 or newer
- Recall is lower than a general-purpose agent, per the README's own benchmark
- Benchmark figures are the authors' own AACR-Bench comparison
- Default mode needs an LLM provider and API key configured first
- no GPU
- Needs Git >= 2.41, LLM provider endpoint (unless delegation mode), Node.js/npm (for npm install)
- Models: Built-in and custom LLM providers, Host agent models in delegation mode
midscene
Midscene.js drives web, Android, iOS, HarmonyOS and desktop UIs from screenshots using a vision model, so tests are written as natural-language steps and assertions instead of selectors. Tests can be YAML cases run by Midscene Test (@midscene/test) or agent calls such as aiAct and aiAssert inside Playwright, Puppeteer, Vitest or Rstest. It produces HTML reports with screenshots and the model's decisions.
Who it is for: QA and frontend engineers writing E2E tests for web, mobile or desktop apps
Strengths
- Locates elements from screenshots, so canvas, icon-only buttons and cross-origin iframes need no selectors
- Same agent API across web, Android, iOS, HarmonyOS and desktop
- Step-level HTML reports with screenshots, plus summary.json for scripts and AI agents
- Works with GPT, Qwen, Doubao, DeepSeek, Gemini, Kimi, including self-hostable models
Weaknesses
- Requires a configured vision-capable model; runs depend on model API calls or self-hosting
- Python and Java SDKs are community projects, not part of the main repo
- Assertions are model judgments, so results can vary between runs and models
- iOS needs WebDriverAgent and Android needs ADB, adding device setup
- Needs Vision-language model API or self-hosted model, Playwright or Puppeteer (web), ADB (Android), WebDriverAgent (iOS), HDC (HarmonyOS)
- Models: GPT, Qwen, Doubao Seed, DeepSeek, Gemini
More in Code review and testing
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.