GitPilot: A CrewAI-based Multi-Agent AI Programming Assistant with Visual Workflow for GitHub Repository Management

GitPilot is a production-grade multi-agent AI programming assistant built on CrewAI and FastAPI, with a React-based management interface. It supports multiple LLM providers including OpenAI, Claude, Watsonx, and Ollama. It can understand the entire codebase structure, present a clear action plan before executing changes, and visualize the thinking process of the multi-agent system.

Most AI programming assistants on the market today are limited to code completion and snippet generation, requiring users to manually copy-paste code into their projects. GitPilot aims to break this limitation by providing an intelligent programming partner that truly understands the project context.

Unlike simple code completion tools, GitPilot's core values include:

• Full Codebase Understanding: Analyzes project structure, file relationships, and cross-repository dependencies
• Plan First: Presents a clear "Answer + Action Plan" before executing any changes
• Multi-Agent Collaboration: Uses CrewAI's multi-agent architecture to divide and handle complex tasks
• Visual Workflow: Intuitively displays the thinking and operation process of the AI system
• Local & Cloud Compatibility: Supports both local file editing and GitHub integration

GitPilot builds a multi-agent collaboration system based on CrewAI. CrewAI is a framework specifically designed for orchestrating multi-agent workflows, allowing AI agents with different roles to collaborate on tasks.

Typical task assignments may include:

• Architect Agent: Analyzes project structure and dependencies
• Developer Agent: Generates and modifies code
• Tester Agent: Writes and runs test cases
• Reviewer Agent: Checks code quality and potential issues

The project uses a modern web tech stack:

• Backend: FastAPI (high-performance Python web framework)
• Frontend: React (modern UI library)
• AI Framework: CrewAI (multi-agent orchestration)
• Deployment: PyPI package distribution, one-click installation via pip

This tech combination balances development efficiency and runtime performance while maintaining good scalability.

Each response from GitPilot is divided into two clear parts:

Answer: Explains what will be done and why in natural language. This helps users understand the AI's decision logic and build trust.

Action Plan: A structured list of steps, clearly marking each file operation:

• 🟢 CREATE – New file
• 🔵 MODIFY – Modify existing file
• 🔴 DELETE – Delete file
• 📖 READ – Read analysis (no changes)

This structure allows users to fully understand all upcoming changes before execution, avoiding unintended modifications.

GitPilot is not tied to a single AI provider and supports seamless switching:

Provider	Supported Models	Features
OpenAI	GPT-4o, GPT-4o-mini, GPT-4-turbo	Strong general capabilities, mature ecosystem
Anthropic	Claude 4.5 Sonnet, Claude 3 Opus	Excellent code understanding and generation
IBM Watsonx	Llama 3.3, Granite 3.x	Enterprise-grade deployment, data privacy
Ollama	Llama3, Mistral, CodeLlama, Phi3	Fully local operation, zero external dependencies

An intelligent model router automatically selects the optimal model based on task complexity, ensuring quality while controlling costs.

GitPilot now supports direct local file system operations, with features similar to Claude Code:

• File Operations: Read, write, search, and delete files within the sandbox workspace
• Terminal Execution: Run shell commands (e.g., npm test, make build, python -m pytest) with timeout and output limits
• Git Integration: Operations like commit, push, diff comparison, staging, merge, etc.
• Security Protection: Path traversal protection to prevent operations from escaping the workspace directory

GitPilot provides advanced session management capabilities:

• Session Recovery: Resume previous work status at any time
• Session Branching: Create branches from the current session to try different solutions
• Checkpoints & Rollback: Create workspace snapshots at key nodes and roll back if necessary
• CI/CD Headless Mode: Run in non-interactive environments via gitpilot run --headless, suitable for automated pipelines