Agent-Gate is a deterministic CI firewall for reviewing AI-generated Pull Requests (PRs), whose goal is to reveal which external PRs need additional context before maintainers read the code. It does not rely on runtime large language model calls, nor does it use AI for judgment; all results are generated based on deterministic rules.
Purpose This tool is used to screen external PRs in public repositories before code review. It presents missing issue links, overly brief descriptions, skipped templates, and excessively large changes, helping maintainers prioritize their review queue based on deterministic facts. First-time contributions are shown only as background information and are never used for scoring. MergeWarden will not call contributions spam, low quality, or AI-generated, nor will it close, label, score, or comment on any PR.
Key capabilities The web interface allows pasting a public GitHub repository, reads summaries of the 30 most recent open PRs, and after removing trusted repository roles, base branches, and known maintenance automation, loads details for at most ten external PRs. The queue is sorted by the deterministic facts a maintainer would otherwise check before reading the code. Detailed scanning accepts a full PR URL or the "owner/repository#number" format, and each scan checks four safety boundaries. The configuration reference lists each deterministic rule and its severity. A CLI is also provided to run the same scan from the terminal without cloning the target repository; the GitHub Action automates detailed PR risk checks but does not sort the review queue.
Usage Simply paste a public repository or full PR URL into the web interface to use it, with no login required. Create a .github/workflows/mergewarden.yml file to enable the Action; for immutable installations, pinning the release commit is recommended, and MergeWarden does not publish or recommend a mutable v0 tag. For larger-scale authenticated queues, use the CLI.
Use cases Suitable for open source project maintainers reviewing external Pull Requests, especially when they need to quickly identify cases that need additional context, carry safety boundary risks, or involve workflow changes. Also suitable for deterministic risk checks on AI-generated PRs through CI automation.
Limitations The web app communicates directly with the public GitHub API and has no telemetry; the queue reads metadata and templates at each PR's exact base commit, and detailed scanning reads only the files required by deterministic rules. It does not perform checkouts and does not execute PR code in the web app or Action; policy comes from the exact base commit, not from the untrusted PR head; analysis never calls language models; incomplete evidence is reported as incomplete and is never presented as passing. Browser-side sample limits are explicit: when 30 summaries are reached or external PRs exceed the ten-item detail limit, a complete queue will not be produced. This tool is experimental, review time savings and maintainer reuse have not yet been proven, and product expansion is paused. No pricing or fee information is provided. Each new rule requires a passing fixture, a failing fixture, and a report snapshot, as detailed in the contribution guidelines. For more information, refer to the documentation index for configuration, Action, CLI, coding tool integration, reproducible research, and the MCP server.files. It does not perform checkouts and does not execute PR code in the web app or Action; policy comes from the exact base commit, not from the untrusted PR head; analysis never calls language models; incomplete evidence is reported as incomplete and is never presented as passing. Browser-side sample limits are explicit: when 30 summaries are reached or external PRs exceed the ten-item detail limit, a complete queue will not be produced. This tool is experimental, review time savings and maintainer reuse have not yet been proven, and product expansion is paused. No pricing or fee information is provided. Each new rule requires a passing fixture, a failing fixture, and a report snapshot, as detailed in the contribution guidelines. For more information, refer to the documentation index for configuration, Action, CLI, coding tool integration, reproducible research, and the MCP server.
