🚀 agentic-cicd: Automate GitLab CI/CD with Natural Language + AI
“Automate your GitLab CI/CD workflows using natural language — powered by AI and Google Cloud.”
✨ Overview
agentic-cicd is an intelligent DevOps agent that allows users to automate complex GitLab CI/CD workflows simply by describing what they want in natural language.
It bridges natural language commands and real infrastructure changes by using an AI-driven multi-agent system that understands, generates, pushes, and monitors CI/CD pipelines on GitLab.
Whether you're a developer looking to set up a CI/CD pipeline or an ops engineer handling deployments — this tool simplifies GitLab automation through LLMs + agents + MCP tool servers.
📟️ Demo
🧠 Key Features
-
🔐 GitLab OAuth Integration – Connect your GitLab account securely.
-
💬 Natural Language Input – Describe what you want, like:
Set up CI to run tests on push and deploy to GKE when main branch is updated. -
🤖 Multi-Agent AI Engine – Understands the request and plans accordingly.
-
⚙️ Pipeline & Infra Generation – Generates
.gitlab-ci.yml, Dockerfile, Terraform, Helm config, etc. -
↻ Push + Merge Request – Pushes the generated files to your GitLab repo and opens an MR.
-
🔍 Pipeline Monitoring – Optionally track status of the build, test, and deployment stages.
-
🧹 MCP Server/Client Architecture – Modular tool abstraction to interface with GitLab API via agents.
🧹 Architecture
↻ System Flow
🧠 Multi-Agent Architecture – LangGraphJS Supervisor Pattern
This project follows the LangGraphJS Multi-Agent Supervisor Pattern to automate GitLab CI/CD workflows using natural language commands.
🧑⚖️ 1. Supervisor Agent
- The central decision-maker.
- Receives natural language input from the user.
- Maintains shared memory and execution state across agents.
- Dynamically delegates to sub-agents using a graph-based control flow.
🔄 Agent Workflow Graph
The Supervisor uses LangGraph to define a directed graph of agents, where edges represent conditional flows based on task results.
🤖 Sub-Agents
Each agent handles a distinct responsibility within the pipeline.
🧭 PlanAgent
“Understand what needs to be done.”
- Parses user intent and decomposes it into logical subtasks.
- Understands GitLab CI/CD terminology and translates user commands accordingly.
- Outputs a plan with goals such as:
- Generate
.gitlab-ci.yml - Create
Dockerfile - Define cloud deployment strategy (e.g., GKE)
- Generate
🧑💻 CodeAgent
“Write the code/configuration needed.”
- Converts the structured plan into CI/CD artifacts such as:
.gitlab-ci.ymlDockerfileTerraform/Helmconfigurations
- Adapts generation based on selected cloud providers (GCP, AWS, etc.)
- Can scaffold standard GitLab CI templates dynamically.
🧪 TestAgent
“Validate the configuration before use.”
- Runs tests on generated configuration to ensure correctness.
- Performs:
- YAML linting / CI syntax validation
- Dry-run of GitLab CI if applicable
- Optional LLM-assisted analysis of pipeline logic
💬 ChatAgent
“Explain what’s happening or answer user queries.”
- Maintains conversation with the user.
- Can answer questions like:
- "What is this pipeline doing?"
- "Why did my build fail?"
- Pushes updates about:
- MR creation
- Pipeline execution
- Deployment status
- Can re-prompt for missing details if required
🔧 GitLab MCP Tool Server (Tool Use via API)
Agents like CodeAgent and ChatAgent interact with GitLab via a dedicated Tool Server (GitLab MCP Server).
🛠 MCP GitLab Tool Server
- A REST API wrapper around GitLab’s official APIs.
- Allows fine-grained agent control over GitLab actions like:
- Getting project info
- Creating branches/commits
- Listing and triggering pipelines
- Monitoring job statuses
- Opening Merge Requests
🤝 MCP Client
- Each agent uses an MCP Client to communicate with the Tool Server.
- Benefits:
- Keeps logic clean and separated from direct API calls.
- Enables observability, tracing, and debugging of agent decisions.
- Makes actions reversible and replicable.
🔄 Feedback Loop
-
User Input
“Set up CI to run tests on push and deploy to GKE when main is updated.”
-
Supervisor Flow
PlanAgent→ Understands the request and breaks into subgoals.CodeAgent→ Generates.gitlab-ci.yml,Dockerfile, etc.TestAgent→ Validates the generated files.GitLab MCP Server→ Applies the changes via GitLab API (opens MR).ChatAgent→ Informs the user and monitors progress.
-
Pipeline Execution
- GitLab CI is triggered.
- Status is periodically updated to the user.
✅ Why LangGraphJS?
-
Structured Reasoning
Flow is modeled as a directed graph for better traceability and control. -
Shared Memory Across Agents
All agents work with a centralized state context. -
Retries and Fallbacks
Automatically retry failed tasks or re-ask user when needed. -
Plug-and-Play Agents
Each agent is modular, testable, and can be extended independently.
✨ The result is a natural-language-driven GitLab CI/CD automation system powered by intelligent agents, tailored infrastructure, and real-time feedback – all orchestrated via LangGraph.
🔐 Authorization:
- Uses the user’s OAuth token to act on their behalf.
- Each API call is scoped by the GitLab OAuth access token.
💽 Tech Stack
| Layer | Tech |
|---|---|
| Frontend | React, React Router, Vite, Tailwind |
| Backend | Node.js, Express, Supabase |
| AI Agents | LangGraphJS, LangChainJS, Gemini/GPT |
| Auth | GitLab OAuth 2.0 |
| Infra Tools | Docker, Terraform, Helm (generated) |
| Hosting | Vercel (frontend), Render (backend) |
🚀 Getting Started
1. Clone the Repository
git clone https://github.com/your-org/agentic-cicd.git
cd agentic-cicd
2. Frontend Setup
cd frontend
npm install
Create a .env file:
VITE_BACKEND_URL=https://your-backend-url.onrender.com
VITE_GITLAB_CLIENT_ID=your_gitlab_client_id
VITE_REDIRECT_URI=https://your-frontend-url.vercel.app/oauth/callback
npm run dev
3. Backend Setup
cd backend
npm install
.env file:
GITLAB_CLIENT_ID=your_gitlab_client_id
GITLAB_CLIENT_SECRET=your_gitlab_client_secret
GITLAB_REDIRECT_URI=https://your-frontend-url.vercel.app/oauth/callback
SUPABASE_URL=your_supabase_url
SUPABASE_KEY=your_supabase_key
GOOGLE_API_KEY=your_gemini_api_key
TAVUS_API_KEY=your_tavus_key
Build and Run:
npm run build
npm run start
🧪 Usage
-
Go to the app homepage.
-
Click "Connect GitLab" and authorize access.
-
Enter your command like:
Set up CI to run tests on push and deploy to GKE when main branch is updated. -
Agent will:
- Understand your intent
- Generate
.gitlab-ci.ymland Dockerfile - Push to a new branch
- Open a merge request
- Trigger a pipeline
-
Watch pipeline logs via GitLab or frontend.
☁️ Deployment Guide
Frontend (Vercel)
- Link GitHub Repo → Vercel Dashboard
- Set environment variables in Settings
- Deploy
Backend (Render)
- Create new Web Service → Connect to GitHub
- Build Command:
npm run build - Start Command:
node dist/index.js - Add your environment variables
❗ Troubleshooting
| Issue | Fix |
|---|---|
CORS error | Ensure backend sets correct Access-Control-Allow-Origin |
OAuth redirect 404 | Check GitLab app redirect URI matches deployed frontend |
Failed to fetch | Backend must be live + CORS enabled |
503 MCP Init | GitLab token expired or MCP endpoint down |
CI config invalid | Re-run or correct command syntax |
🤝 Contributing
We welcome PRs and ideas!
# Fork -> Clone -> Create Branch -> Code -> Commit -> PR
Please open issues for:
- Bugs
- New agent tasks
- GitLab features you'd like automated
📄 License
MIT License © 2025 MemeVerse CICD Dev Team
💡 Inspired By
🗂️ Folder Structure (Simplified)
agentic-cicd/
│
├── frontend/
│ ├── App.jsx
│ ├── pages/
│ └── components/
│
├── backend/
│ ├── agents/
│ ├── routes/
│ |── services/
|-─ mcp-server/
│
├── .env.example
└── README.md
📬 Contact
For inquiries or collaborations: 📧 shalwinsanju.25cs@licet.ac.in 🐦 Twitter: @samshalwin