MR Impact Radar

MR Impact Radar

Inspiration

Every merge request carries risk. In large codebases, it is not always obvious which files could be affected, who should review the changes, or whether a seemingly simple update could break production. We built MR Impact Radar to predict the hidden “blast radius” before code gets merged.

What it does

MR Impact Radar is a multi-agent AI system powered by Gemini 3 Flash that analyzes GitLab merge requests and helps reviewers understand risk before merging. It provides:

Risk Score (0–100) — A quantified assessment of merge risk
Blast Radius — A visual map of affected files and dependencies
Smart Reviewers — Suggested reviewers based on code ownership and affected areas
Review Checklists — AI-generated items for reviewers to verify
Code Explanations — File-level explanations for why changes are marked HIGH, MEDIUM, or LOW risk

How we built it

We built a 3-agent pipeline using Gemini AI:

Diff Agent — Analyzes code changes and explains what changed
Dependency Agent — Maps affected files, dependencies, and reviewer suggestions
Risk Agent — Scores the merge risk and generates review recommendations Merge request data is fetched through GitLab MCP using the Model Context Protocol, with the GitLab REST API as a fallback for reliability and line-level availability.

Challenges we ran into

One major challenge was GitLab MCP authentication. GitLab MCP requires OAuth tokens with the mcp scope, which must be obtained through browser-based Dynamic Client Registration. Because of this limitation, deploying directly to Cloud Run was not suitable for our use case. We deployed the application to a Compute Engine VM instead, which gave us more control over the authentication flow and runtime environment.