CoderX: A Local-First, Open-Source AI-Powered Website Builder with Multi-Agent Code Generation

Abstract

Software development workflows have been disrupted by the arrival of large language model (LLM)-based code generation tools. However, existing platforms such as Lovable and Vercel v0 transmit user prompts and source code to remote cloud servers, raising concerns around data privacy, intellectual property, and offline availability. This paper presents CoderX, a locally hosted, open-source AI-powered website builder that transforms plain English descriptions into complete, deployment-ready full-stack web applications while running entirely on the developer\'s own machine. CoderX employs a novel coderxArtifact streaming parser that interprets structured XML action tags emitted by local LLMs in real time, writing each generated file to disk character by character in a Monaco code editor. A coordinated multi-agent crew pipeline comprising Planning, Backend, Frontend, and QA crews with a total of sixteen specialist worker agents manages the full generation lifecycle from intent analysis through to automated dependency installation and live preview launch. A GitHub Intelligence Agent mines open-source repositories before generation begins, grounding output in proven real-world code patterns. Empirical evaluation demonstrates a 49/49 unit test pass rate, zero TypeScript compilation errors, and generation of complete nine-component React applications in under three minutes using the Qwen2.5-Coder-7B model at Q4 quantisation on consumer hardware. The project is released under the MIT licence with full contributor community infrastructure including CONTRIBUTING.md, issue templates, and automated validation workflows.

Introduction

Software development requires expertise across multiple layers, including database design, backend APIs, frontend components, testing, and deployment. Although AI-powered coding tools have significantly accelerated development, most existing solutions rely on cloud-based large language models (LLMs) and generate only isolated code snippets rather than complete, integrated applications.

This paper introduces CoderX, an open-source, locally hosted AI website builder designed to overcome these limitations. CoderX performs all code generation on the developer’s own machine using the Ollama runtime, ensuring privacy, data sovereignty, and offline functionality. It employs a multi-agent architecture consisting of eight specialized agents coordinated by an orchestrator to generate fully integrated full-stack TypeScript applications.

Problem Statement

Existing AI code-generation tools face several challenges:

1. Cloud Dependency – User prompts and code are sent to remote servers, raising privacy and security concerns.
2. Partial Generation – Tools generate separate components instead of complete, interconnected applications.
3. Opaque Execution – Developers cannot observe or control the generation process.
4. Provider Lock-In – Many platforms require subscriptions or specific cloud-based AI services.
5. Ephemeral State – Browser-based tools may lose project data when sessions end.

Gap Analysis

The study identifies several shortcomings in current AI-assisted development systems:

• Existing multi-agent frameworks generate better software but depend on cloud-hosted LLMs.
• No current system combines local LLM inference with multi-agent full-stack generation.
• Real-time file streaming during code generation is largely restricted to cloud platforms.
• Existing tools do not use GitHub repository mining before generation to learn current coding practices.
• Most tools require manual steps to install dependencies, run applications, and preview results.

Literature Survey Findings

Previous systems provide valuable contributions but have limitations:

• Multi-agent frameworks improve software quality but rely on cloud inference.
• Code-generation models such as Codex and Copilot focus mainly on function-level assistance.
• Platforms like Bolt.new provide real-time generation but require cloud infrastructure.
• Local runtimes such as Ollama support offline LLMs but lack development workflow automation.

Existing Systems

Three major categories of tools were reviewed:

• Cloud-based platforms (e.g., Lovable and Bolt.new) provide full-stack generation but require internet connectivity and cloud processing.
• AI-enhanced IDEs (e.g., Cursor and GitHub Copilot) assist developers but do not generate complete applications automatically.
• Local LLM runtimes (e.g., Ollama) support offline AI inference but lack integrated project generation capabilities.

Proposed CoderX System

1. Local-First Architecture

CoderX prioritizes:

• Complete local execution
• Privacy and security
• Transparent agent actions
• Fully integrated application output

2. Streaming Code Generation

A custom coderxArtifact parser processes structured XML-based outputs from the LLM. As files are generated, code appears in real time within the editor and is automatically written to disk, creating an interactive development experience.

3. Multi-Agent Crew Architecture

The generation workflow is managed by specialized agent teams:

• Planning Crew: Defines entities, routes, packages, and project structure.
• Backend Crew: Creates schemas, services, APIs, and tests.
• Frontend Crew: Develops components, pages, hooks, and design elements.
• QA Crew: Performs type checking, testing, security scanning, and build validation.
• Reviewer Agent: Ensures consistency across frontend and backend layers.
• Documenter Agent: Generates project documentation and configuration files.

4. GitHub Intelligence Agent

Before generation begins, a dedicated agent analyzes relevant public GitHub repositories to:

• Identify common architectures and dependencies.
• Extract coding patterns and best practices.
• Provide contextual guidance to other agents.

Conclusion

This paper presented CoderX, a locally hosted, open-source AI-powered website builder that generates complete full-stack web applications from natural language descriptions using a coordinated multi-agent crew pipeline and real-time streaming file writes. The system demonstrates that privacy-preserving, fully offline AI application generation is practically achievable on consumer hardware using quantised open-weight models through the Ollama runtime. The coderxArtifact streaming parser, adapted and extended from the concepts introduced in Bolt.new, enables a transparent generation experience where every file materialises in the Monaco editor as the LLM produces it. The four-crew, sixteen-worker agent architecture produces integrated TypeScript codebases that consistently compile cleanly and pass automated tests without manual intervention. The 49/49 unit test pass rate, zero TypeScript compilation errors, and 100% npm install and live preview success rate across twenty evaluation sessions confirm that the system delivers reliable, production-quality output at a scale and speed that meaningfully reduces development time. CoderX is released under the MIT licence with complete contributor community infrastructure, and its modular architecture invites community participation in extending its capabilities.

References

[1] S. Hong, X. Zheng, J. Chen, Y. Cheng, and C. Wu, \"MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework,\" in Proc. ICLR, 2024. [2] C. Qian, X. Cong, C. Yang, W. Chen, Y. Su, and M. Sun, \"Communicative Agents for Software Development,\" in Proc. ACL, 2024. [3] StackBlitz Inc., \"Bolt.new AI Full-Stack Web Development,\" [Online]. Available: https://bolt.new. [Accessed: Apr. 2026]. [4] M. Chen, J. Tworek, H. Jun et al., \"Evaluating Large Language Models Trained on Code,\" arXiv:2107.03374, 2021. [5] W. Jiang et al., \"Qwen2.5-Coder Technical Report,\" arXiv:2409.12186, 2024. [6] J. S. Park, J. O\'Brien, C. J. Cai, M. R. Morris, and M. S. Bernstein, \"Generative Agents: Interactive Simulacra of Human Behavior,\" in Proc. UIST, 2023. [7] GitHub Inc., \"GitHub Copilot Your AI Pair Programmer,\" [Online]. Available: https://github.com/features/copilot. [Accessed: Apr. 2026]. [8] D. Huang, Q. Bu, J. M. Zhang, and H. Qian, \"AgentCoder: Multi-Agent-based Code Generation with Iterative Testing and Optimisation,\" arXiv:2312.13010. [9] J. Wei, X. Wang, D. Schuurmans et al., \"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,\" in Proc. NeurIPS, 2022. [10] Ollama Inc., \"Ollama Get up and running with large language models,\" [Online]. Available: https://ollama.com. [Accessed: Apr. 2026].

Copyright

Copyright © 2026 T V V N Satyanarayana, T S S D Teja Reddy, V K N S Adeep Reddy, P Vamsi Krishna. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.