Greptile vs Cynical Sally vs Night Market: PR Review (2026)
A. Frans
Published November 16, 2026
Table of Contents
Greptile's skills loop a pull request until a hosted reviewer gives it 5/5. Cynical Sally scores your diff from 0 to 10 and gives you 90 free reviews a month. Claude Night Market bundles 23 plugins with a PR review command, commit-message helpers and safety hooks. All three plug into Claude Code, and all three send something off your machine, which is the part most install guides skip.
This comparison is about that second question as much as the first. We read each project's README and the skill files themselves on November 16, 2026; star counts come from the GitHub API that day. We didn't run the three against a shared test repo, so treat the quality notes as what the docs promise, not a benchmark.
If you want the Superpowers review skills (requesting-code-review, receiving-code-review, finishing-a-development-branch), we covered them in our code review skills roundup. They run entirely inside your Claude Code session, so they're the baseline the three tools below get measured against.
| Greptile skills | Cynical Sally | Claude Night Market | |
|---|---|---|---|
| What it is | 3 agent skills + a plugin with an MCP server | CLI + MCP server | Marketplace of 23 plugins |
| GitHub stars | 435 | 102 | 350 |
| License | MIT | MIT | MIT |
| Review engine | Greptile's hosted reviewer | Sally's server, which calls Anthropic's API | Your Claude Code session, plus optional external model CLIs |
| Free use | Skills are free; Greptile itself is a paid service | 90 reviews/month, no account | Free |
| Where your code goes | Greptile (via your repo integration) | Sally's server, files you name only | Stays local unless delegation or posting hooks fire |
| Best for | Teams already paying for Greptile | Solo devs who want a blunt second opinion | People who want a full workflow kit and will read the config |
Greptile skills: a loop around a reviewer you already pay for
The greptileai/skills repo ships three skills:
check-prlooks at an open PR, MR or Perforce changelist for unresolved comments, failing status checks and a thin description, then offers to fix and resolve them.cli-reviewruns a Greptile CLI review against your local checkout and summarizes it.greplooptriggers a Greptile review, fixes every actionable comment, pushes, and re-triggers, repeating until Greptile reports 5/5 confidence and zero unresolved comments.
Install all three globally with the skills CLI:
npx skills add greptileai/skills -g
Or install the plugin, which adds Greptile's MCP server and login and review skills:
/plugin marketplace add greptileai/skills
/plugin install greptile@greptile-plugin
check-pr is the most useful skill for people who don't use Greptile at all. It only needs git plus an authenticated gh (GitHub), glab (GitLab) or p4 (Perforce), so it works on any PR with human reviewers. greploop needs Greptile installed on the repository; without it there's nothing to loop against.
The security story here is good. Each skill's frontmatter limits it to Bash(gh:) Bash(glab:) Bash(git:) Bash(p4:), so Claude can't wander off and run arbitrary shell commands under that skill. Your code reaches Greptile through the integration your team already approved, not through anything new the skill adds.
One warning about greploop. "Loop until the reviewer has zero comments" optimizes for the reviewer, and an agent that fixes every comment will also fix the wrong ones. Pair it with the Superpowers receiving-code-review skill, which tells Claude to verify each suggestion against the codebase and push back with reasons instead of agreeing.
Cynical Sally: a blunt reviewer with a dry-run button
Cynical Sally is technically a CLI and an MCP server, not a SKILL.md skill, but it's used the same way. It scores code from 0 to 10, lists issues with evidence, and suggests fixes, in a deliberately harsh voice.
npm install -g @cynicalsally/cli
sally roast --staged # review what you're about to commit
sally roast --diff main # review the branch against main
sally roast ./src/ -m full_truth
To use it inside Claude Code:
claude mcp add cynical-sally -- npx @cynicalsally/cli mcp
It needs Node.js 18 or newer. The free allowance is 90 reviews a month with no account; reviews are tied to a random device ID until you link an email.
This one sends code off your machine, and the README is unusually clear about it. Files you name go over HTTPS to Sally's server, which runs the analysis through Anthropic's API. The project says source is processed in memory, never written to disk or logs, and never used for training. You can check what would be sent first:
sally roast --dry-run ./src/
That prints every file path, size, token estimate and a SHA-256 receipt, then sends nothing. Secret files like .env and key files are skipped locally before upload. We'd still keep it off client code covered by an NDA unless your contract allows third-party processors.
For CI, sally roast ./src/ --fail-under=5 --ci fails the build when the score drops below 5. That's a blunt gate, so start with a low threshold and raise it once you trust the scores. The repo's last push was July 4, 2026, which makes it the least recently updated of the three.
Claude Night Market: a whole workflow, with defaults to check
Claude Night Market is a plugin marketplace. Its own description lists 23 plugins, 186 skills, 128 commands and 54 agents. The two that matter for PRs are sanctum (git workflows, commits, PR prep) and pensive (code review):
/plugin marketplace add athola/claude-night-market
/plugin install sanctum@claude-night-market
/plugin install pensive@claude-night-market
It needs Claude Code 2.1.16 or later and Python 3.9 or later for its hooks. After installing, run claude --init-only once.
sanctum gives you /prepare-pr, /pr-review, /fix-pr, /resolve-threads, /commit-msg and /validate-pr. pensive gives you /full-review, which picks from specialist reviewers for APIs, architecture, bugs, tests, Rust, shell scripts and performance, plus a blast-radius skill that scores the risk of a change.
The safety features are the best of the three. Hooks block rm -rf and git push --force, refuse to read .env, .netrc and SSH keys, and strip invisible characters and prompt-injection phrases from fetched web content.
Read the "Network and data" section of the README before you install, because three parts reach off your machine:
1. A star prompt at session start checks through gh whether you've starred the repo. It asks once per session and never stars without consent. Set CLAUDE_NIGHT_MARKET_NO_STAR_PROMPT=1 to turn it off. 2. Learnings posting is on by default. When ~/.claude/skills/LEARNINGS.md has content, a stop hook posts a skill-usage discussion through gh, to the repo set in ~/.claude/skills/discussions/config.json or, failing that, the current repo. Set auto_post_learnings: false in that file if you work in private or client repos. 3. Delegation, also on by default, sends the prompt and the file contents it carries to the first external model CLI that answers, trying Gemini, Qwen, MiniMax, GLM, Muse, Codex and OpenCode in that order. Set CONJURE_DELEGATION=off for one run, or "enabled": false in ~/.claude/hooks/delegation/config.json to stop it on that machine.
None of this is hidden; the README spells it out. But "on by default" plus "posts to the current repo" is a combination worth turning off on day one if you don't want it.
Which one to install
- Your team pays for Greptile: install
greptileai/skills. Usecheck-prdaily andgreploopon PRs you want clean before a human looks. - You work alone and want a second opinion before pushing: Cynical Sally, run with
--dry-runthe first time. - You want commit, PR and review commands in one kit: Night Market's
sanctumandpensive, after you've set the two config flags above. - Your code can't leave the machine at all: skip all three and use the Superpowers review skills, which dispatch a reviewer subagent inside your own session.
Whatever you pick, read the skill files before installing. Our guide to auditing an agent skill before you install it covers what to look for, and our full list for developers covers hosted review tools such as CodeRabbit if you'd rather not run review inside your agent at all.
FAQ
Do the Greptile skills work without a Greptile subscription?
Partly. check-pr only needs git and an authenticated gh, glab or p4 CLI, so it works on any PR. greploop and cli-review need Greptile itself, which is a paid hosted service.
Does Cynical Sally upload my code?
Yes, the files you point it at are sent over HTTPS to its server, which uses Anthropic's API. The project says code isn't stored or trained on. Run sally roast --dry-run to see the exact payload first.
Is Claude Night Market safe to install?
It has stronger guardrails than most plugin packs, including blocks on force-push and credential reads. Two network features are on by default: learnings posting and delegation to external model CLIs. Turn both off in config if you work on private code.
Can these tools replace human code review?
No. They catch missing tests, failing checks and obvious bugs before a human looks, which saves reviewer time. Design decisions and whether the change should exist at all still need a person.
Which is best for GitLab or Perforce?
Greptile's check-pr and greploop detect GitHub, GitLab and Perforce automatically, and they're the only tools in this comparison that document Perforce support.
Share this article
📄Related Articles
Get More AI Tool Guides
New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.