Maven

Your GitLab project never sleeps. Neither does Maven.

Comment 1

Inspiration

We were inspired by the AI Paradox: while AI can generate code faster than ever, the surrounding processes—security reviews, CI/CD, issue triage—lag behind. We wanted to create a system that doesn't wait for prompts but acts autonomously to maintain code quality and workflow efficiency.

What it does

Maven is an autonomous DevSecOps swarm inside your GitLab environment. It monitors pipelines, merge requests, issues, and security scans in real time. When events occur—like a failing pipeline or new vulnerability—Maven diagnoses, fixes, and executes actions automatically. It features a Queen Agent (powered by Claude) orchestrating specialized worker agents for security, testing, compliance, deployment, and triage, all self-healing and adaptive.

How we built it

We built Maven on top of the open-source Hive agent framework by Aden, specialized for GitLab. The Queen Agent (Claude via Anthropic API) generates dynamic worker graphs for each event, while the GitLab Duo Agent Platform executes tasks. We integrated LiteLLM for model-agnostic routing, MCP for tool access, and a WebSocket-based observability dashboard for live tracking.

Challenges we ran into

  • Designing a self-healing system where workers retry intelligently without human intervention.
  • Dynamically generating worker graphs that adapt to different project languages and structures.
  • Ensuring secure credential storage and API access across multiple agents.
  • Balancing model usage for cost-efficiency with responsiveness.

Accomplishments that we're proud of

  • Fully autonomous response to GitLab events with measurable impact on workflow efficiency.
  • Self-healing loops that retry and adapt after failures.
  • Real-time observability dashboard tracking success rates, costs, and agent activity.
  • Green agent routing that balances complexity and cost per task.

What we learned

  • Event-driven autonomy in DevSecOps is feasible and highly effective.
  • Multi-agent orchestration requires careful design of memory, context, and delegation.
  • Human-in-the-loop escalation is crucial when confidence is low, maintaining reliability.
  • Dynamic workflows outperform static pipelines for complex, evolving projects.

What's next for Maven

  • v1.1: Natural language-based custom worker creation.
  • v1.2: Cross-project swarms for GitLab groups.
  • v1.3: Memory and learning per repository for improved accuracy.
  • v2.0: Maven Marketplace for sharing community-built worker templates.

Built with

  • Languages & Frameworks: Python 3.11+, Hive framework
  • Platforms: GitLab, GitLab Duo Agent Platform
  • APIs: Anthropic Claude API, LiteLLM
  • Other Tech: WebSocket dashboard, MCP tool protocol, encrypted credential storage

Try it out

Built With

Share this project:

Updates