Automated code review with large language models is increasingly available through hosted tools such as GitHub Copilot Code Review and Anthropic’s Claude Code. However, these systems are typically built on frontier-scale models and leave two practical questions open: how to make LLM-generated review trustworthy enough to present directly to developers, and how to run automated review on infrastructure that is not paid. This thesis addresses both questions as engineering problems rather than as problems that must wait for stronger models. It presents a multi-agent code-review framework for GitLab merge requests, implemented in TypeScript and executed entirely on Groq’s free-tier endpoint. The framework combines three main architectural choices: a coordinator that routes each rule family to a focused agent through a static lookup table; a default agent model (openai/gpt-oss-20b, 20B parameters) with selective escalation to a 32B model when the output is ambiguous; and a mandatory grounding gate that drops any finding whose evidence is not a literal substring of the diff or whose claimed line is not sufficiently close to a changed line.