local-multi-agent-company: Localized Multi-Agent Software Development Team Architecture

local-multi-agent-company is a localized multi-agent software development team architecture project for the Unraid platform. It uses LangGraph to orchestrate 17 professional Worker roles, supporting GitHub automation, hierarchical deployment, and local LLM routing to achieve a controllable AI-driven software development process. The core of the project is to simulate the division of labor and collaboration of a professional development team, solving the problem that a single AI assistant cannot meet the professional requirements of multiple links, while balancing data privacy, controllability, and development efficiency.

Today, with the popularity of AI-assisted programming, a single AI assistant can hardly meet the professional requirements of multiple links such as requirement analysis, architecture design, coding implementation, and test verification at the same time. The local-multi-agent-company project aims to simulate the operation process of a complete software development company by building a multi-agent system, allowing AI to collaborate like a professional team.

The project includes 17 specialized Workers with clear division of labor:

• Orchestrator: Core orchestrator, responsible for task reception, workflow control, and state persistence;
• Requirement Analysis Layer: requirements-worker extracts requirements, cost-worker estimates resources, human-resources-worker recommends Worker configurations;
• Research & Design Layer: research-worker analyzes repositories, architecture-worker designs architectures, data-worker provides data processing suggestions, ux-worker provides UI suggestions;
• Implementation Layer: coding-worker executes coding, reviewer-worker reviews code, test-worker performs testing, security-worker identifies risks;
• Delivery Layer: validation-worker verifies requirements, documentation-worker creates documents, github-worker manages code submissions, deploy-worker executes hierarchical deployment, qa-worker performs testing, memory-worker persists learning outcomes;
• Additionally, there is a web-ui that provides task management and monitoring interfaces.

The project adopts a local-first design, containerized deployment on the Unraid platform, uses local LLM endpoints (default Mistral and Qwen series), and keeps data local to ensure privacy. In terms of controllability, hierarchical deployment is default (manual approval required from Staging to production), and approval is enforced at key nodes. Orchestration uses the LangGraph framework, supporting complex state transitions and parallel execution; state is persisted to an SQLite database to ensure task recoverability.

The project supports flexible model routing: Mistral is used for lightweight tasks (document generation, classification), Qwen for heavy-load tasks (architecture design, security review), and each Worker can be configured with model parameters independently. It deeply integrates GitHub as the code management source; github-worker is responsible for code submission and PR creation, and uses a repository whitelist and change confirmation mechanism to ensure security.

Security mechanisms include: untrusted external content handling, prompt injection identification, key management (local file storage, Docker read-only mount), and Shell command whitelist. The web interface provides functions such as task management, Worker configuration, and trusted source management; the Debug center supports system snapshot download and runtime file diagnosis, facilitating troubleshooting and configuration optimization.

Applicable scenarios: Local developers/small teams, privacy-sensitive enterprises, multi-agent architecture researchers, Unraid users. Usage recommendations: Start with simple tasks for the first deployment; adjust model routing and timeout configurations according to hardware; use the approval mechanism to maintain supervision at key nodes; regularly optimize learning outcomes accumulated by memory-worker; verify on a small scale first before expanding to production environments.

local-multi-agent-company represents an important attempt in the development of AI-assisted programming towards professionalization and systematization. By simulating the division of labor in real teams, it demonstrates that multi-agent architecture has more reliable and controllable automation capabilities than a single AI. Although deployment and tuning require technical investment, it has important research value for users exploring the boundaries of AI development applications.