Quelltest: the free, offline gate for AI-written tests

AI-generated tests often pass without truly validating the underlying code. Quelltest addresses this problem by deliberately breaking code to verify that tests can actually detect bugs. The tool operates entirely offline and locally, ensuring that no code leaves the developer's machine. The verification process works by creating throwaway copies of the codebase and introducing small, deliberate bugs called mutants. When a test still passes despite these intentional defects, it reveals that the test was never genuinely protecting that code. This approach exposes weak assertions, untested edge cases, and gaps in test coverage that traditional green checks miss. Quelltest provides three primary commands. The "verify" function plants deliberate bugs and confirms tests fail appropriately. The "find" function identifies risky, untested spots in the codebase, prioritizing areas where bugs commonly hide such as guards, fallbacks, and thrown errors. The "gate" function compares tests against a previous version and fails the build when tests have been deleted, skipped, or weakened—preventing developers or AI agents from inadvertently reducing test quality. The tool supports JavaScript and TypeScript through an npm package compatible with Jest, Vitest, Mocha, and node:test on Node.js 20 and above. A Python version is available for Python 3.11 and higher. Both packages include MCP server integration, allowing AI coding agents in Claude Code, Cursor, and other MCP clients to run these checks on their own generated tests. Agents cannot mark work as complete until quelltest validates the tests meet quality standards. Quelltest is free and released under open-source licenses—Apache-2.0 for the npm package and MIT for Python. The tool requires no external services or network connectivity. According to the documentation, it found 11 of 18 known defects reported against a real codebase with zero false alarms on seven clean files, and rejected seven of seven weak tests at the recommended 60% threshold.