Review-Gate: keeping one Cursor request open until you say it is done
Review-Gate V2 is a powerful rule for the Cursor IDE that helps you get up to 5x more value from your monthly requests. It creates an interactive loop where the AI waits for your follow-up commands—via text, voice, or image upload—allowing you to perform deep, iterative work all within a single request.
At a glance
- What is it?
- LakshmanTurlapati/Review-Gate is a Cursor rule that makes the agent stop and ask before it signs off on a request, so follow-ups happen inside the original request instead of starting new ones. V2 does this through an MCP tool that opens a popup accepting text, voice and images.
- Who is it for?
- Review-Gate fits a Cursor user on a request-counted plan who does long, iterative work and wants corrections inside one request instead of five. It is the wrong tool on a usage-metered plan, where spending more tool calls per request saves nothing, and for anyone who cannot install an MCP server and an extension on their machine.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 169 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 18, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The request budget problem it attacks
The README opens with a specific complaint rather than a mission statement. Give Cursor a complex task and it might spend about five of the roughly 25 tool calls available to that request, then stop. Every small correction after that starts a new request, and the author's plan allows about 500 requests per month, so corrections are what burns the quota.
Review-Gate is built to invert that. The agent is told the request is not finished until the user says so, and the follow-ups land inside the original request's lifecycle and tool call budget. The README's headline is up to five times the value from the monthly allowance, and it immediately hedges its own number: maybe not always a perfect five times, but you get the idea.
How the V2 loop runs
The mechanism is an MCP tool named review_gate_chat. The sequence the README describes runs like this: you give Cursor a task, which counts as one main request. The agent does its work. Instead of finishing, it calls review_gate_chat, and a popup appears inside Cursor with text, microphone and image input, while the agent announces in chat that it is waiting.
You then type a follow-up, speak one, or drop in a screenshot or mockup. The agent reads that input, does more work drawing on the original tool call budget, answers in the main chat, and opens the popup again. The loop continues until you type TASK_COMPLETE in the popup.
That last detail is the one to remember, because it is the only exit. The popup reopens after every response, so a session that never receives TASK_COMPLETE never closes.
Installing Review-Gate V2
The README sends readers to V2/INSTALLATION.md for the canonical file list, which it says covers V2/review-gate-v2-2.7.3.vsix, V2/ReviewGateV2.mdc, a manifest-backed packaging helper and the merge-safe MCP setup. For the normal path, the two steps are clone and run the installer.
On macOS:
git clone https://github.com/LakshmanTurlapati/Review-Gate.git
cd Review-Gate/V2
./install.shOn Windows:
git clone https://github.com/LakshmanTurlapati/Review-Gate.git
cd Review-Gate/V2
./install.ps1The README notes the PowerShell installer may need administrator privileges. When it finishes, the MCP server should be available to Cursor and the extension should be installed, so the next complex task should end in a popup rather than a finished answer.
Maintainers rebuilding the artifact get their own path. The v2.7.3 release notes say npm run package now goes through a portable wrapper so it works on environments that expose python3 but no python alias, and that V2/review-gate-v2-2.7.3.vsix is treated as the only supported packaged artifact.
What the installer changes on your machine
Four things, per the README. It installs dependencies through Homebrew or Chocolatey, including SoX for audio capture and some Python packages. It installs the MCP server globally with its configuration. It installs the Cursor extension that renders the popup. And it edits your MCP configuration by merging only the review-gate-v2 entry, so existing MCP servers are left alone.
That last point is the one worth checking afterwards. Any tool that edits a shared MCP config file can break the other entries, and the merge-only claim is the project's answer to that. The repository also carries scripts/ and tests/ directories, and a .github/ workflow added in 2.7.3 that validates the release manifest, runs the release-surface regression suite, builds the VSIX and uploads it as a workflow artifact.
V1 against V2
The README includes a comparison table, and the differences are concrete. V1 was a terminal-based Python script taking text only, on macOS and Linux, set up by hand. V2 is a native MCP tool with a popup interface, text plus voice plus images, on macOS, Windows and Linux, installed by a one-click script.
Speech-to-text is the sharpest change: V1 had none, V2 runs local Faster-Whisper, which means transcription happens on your machine rather than through an API. Image upload and a real-time MCP status indicator are new as well. The README is careful about one thing though. It says Whisper speech-to-text works flawlessly on macOS and is implemented for Windows and Linux, which is not the same claim for all three.
The repository still carries ReviewGate.mdc at the root, the V1 rule, alongside V2/.
Where the five times claim stops holding
The multiplication only means something if requests are your scarce unit. The whole design spends more tool calls inside one request to avoid spending another request. On a plan where usage is metered by tokens or by compute rather than by request count, the mechanism saves nothing and costs extra turns.
The five times figure also carries no measurement behind it in the README. It is the author's framing of a quota arithmetic, hedged in the same sentence that introduces it, and no before-and-after request counts are published.
Then there is the surface area. Running this means an MCP server, a Cursor extension, SoX, and Python packages on your machine, installed by a script that may ask for administrator rights on Windows. Anyone who has to clear that with a security team should read install.ps1 first. And the loop needs a human in it: an agent left waiting on a popup that never receives TASK_COMPLETE is a request that never closes.
The V1 rule as a lighter option
ReviewGate.mdc at the repository root is the V1 approach, and it remains a real alternative: a plain Cursor rule file that tells the agent to check in before it finishes, with no MCP server, no extension, no installer and no audio stack. Text only, and the README lists it as macOS and Linux.
Choosing it means giving up the popup, the voice input and the image upload, and doing the setup by hand. What you keep is the core behaviour, which is the instruction to wait for a go-ahead. If your objection to V2 is the installed surface area rather than the missing modalities, the rule file alone is the smaller risk.
Release state and licence
The latest release is v2.7.3, published 2026-04-02, and the last push to the repository was on 2026-04-02. That release added the GitHub Actions workflow and made V2/review-gate-v2-2.7.3.vsix the single canonical artifact, with the older 2.6.x VSIX files removed in an earlier version.
On licensing the repository contradicts itself. The README carries an MIT badge, but GitHub reports no licence for the repository and the file listing shows no LICENSE file. MIT is what the project claims, not what the repository provides, so anyone reusing the code should ask the author to add the file before treating the badge as settled.
Upkeep cost is mostly re-installation. A Cursor update that changes how MCP servers or extensions are loaded can break the popup, and the fix path is to re-run the installer.
Editorial conclusion
Review-Gate fits a Cursor user on a request-counted plan who does long, iterative work and wants corrections inside one request instead of five. It is the wrong tool on a usage-metered plan, where spending more tool calls per request saves nothing, and for anyone who cannot install an MCP server and an extension on their machine. Before relying on it, read V2/INSTALLATION.md for the canonical file list and check your MCP config after the installer runs, since that is the file it edits.
Frequently asked questions
Does Review-Gate work on Windows?
Yes. The README gives a PowerShell installer, ./install.ps1, run from the Review-Gate/V2 directory, and notes it may need administrator privileges. The platform badge lists macOS, Windows and Linux.
How do I end the Review-Gate loop?
Type TASK_COMPLETE in the popup. Until you do, the agent reopens the popup after each response, so the request stays open.
Which file is the canonical V2 release?
The README and the v2.7.3 release notes both name V2/review-gate-v2-2.7.3.vsix as the only supported packaged artifact, with the older 2.6.x VSIX files removed.
Does the voice input send audio to a cloud service?
No. The README says transcription runs on local Faster-Whisper. The installer does add SoX for audio capture, through Homebrew on macOS or Chocolatey on Windows.
Community notes