Your AI engineering manager for GitLab

Warden watches your repository, surfaces concrete risks with evidence, and proposes fixes — but never mutates your project without your approval. Built for the Rapid Agent hackathon.

The problem

Repos accumulate debt. Nobody files the issues.

Missing tests, oversized files, TODOs that never become tickets, broken pipelines, commits that land on main without review — someone is supposed to notice. Warden does.

How it works

Observe → Reason → Propose → Approve → Execute

A human-in-the-loop command bus where the agent watches and reasons, and only a deterministic executor creates GitLab issues after you click Approve.

Connect GitLab

Run a scan

Static analysis, GitLab MCP context, and Gemini reasoning run in one pass.

Review findings

Prioritized issues with severity scores and file-level evidence — not vague AI prose.

Approve proposals

You stay in control. Warden proposes fixes; nothing writes to GitLab without you.

Issue created

Approve once and Warden creates a real GitLab issue via the REST API.

What Warden detects

Five focused finding categories

Each finding includes severity, a priority score, and evidence references so results are auditable — not vague AI prose.

MISSING_TESTS

Missing Tests

Surfaces files without corresponding test coverage.

MAINTAINABILITY

Maintainability

Flags oversized files and structural complexity risks.

TECHNICAL_DEBT

Technical Debt

Tracks TODOs, FIXMEs, and debt that never becomes tickets.

RISKY_CHANGE

Risky Changes

Detects direct commits and changes that bypass review.

CI_CD

CI / CD

Monitors pipeline health and broken automation patterns.

Built different

Trust by design

Human-in-the-loop

The agent observes and reasons. A separate executor creates issues only after you approve.

Agent Builder + Gemini

Google Cloud Agent Builder runs structured reasoning over your repository context.

GitLab MCP integration

Merge requests, pipelines, and commits are gathered through the GitLab MCP server.

Architecture

Built for real scans, not chat demos

Web on Vercel

OAuth login, repo selection, scan enqueue, and proposal approval UI.

Worker on Cloud Run

Checkout, static analysis, GitLab MCP context, and Agent Builder reasoning.

Neon Postgres

Users, repos, scans, findings, proposals, and a full activity audit trail.

WARDEN

Start watching your repository

Connect GitLab, run your first scan, and approve a proposal to create a real issue — in minutes.