Product Introduction
Definition: Mozzie is an open-source, local-first AI agent orchestrator designed as a desktop application. It functions as a sophisticated management layer and workspace for autonomous AI coding agents, built using the Tauri 2.0 framework with a Rust-based backend and a React-based frontend.
Core Value Proposition: Mozzie exists to solve the "sequential bottleneck" in AI-assisted development. By providing a centralized environment to manage multiple AI agents—such as Claude Code, Gemini CLI, and Codex—in parallel, it allows developers to scale their productivity without cloud dependencies. The primary value lies in its ability to break complex projects into discrete work items, assign them to different agents, and manage the resulting git lifecycle through isolated worktrees, ensuring that AI-driven code generation is organized, concurrent, and secure.
Main Features
Multi-Agent Parallel Orchestration: Mozzie allows for the simultaneous execution of multiple AI coding agents. Each agent runs in its own isolated terminal environment within the desktop app. It supports the Agent Communication Protocol (ACP) over stdio for deep integration with tools like Claude Code, while maintaining a fallback mechanism for standard CLI-based agents. This allows a single developer to act as a manager for a "swarm" of agents working on different parts of a codebase at once.
Git Worktree Isolation & Dependency Graphing: To prevent file system conflicts when multiple agents work on the same repository, Mozzie utilizes native Git worktrees. Every work item is assigned its own temporary directory and branch. The system features a built-in dependency management engine with cycle detection; work items can be "blocked" by others, automatically launching once their dependencies are resolved. This enables complex, multi-stage refactoring or feature development where tasks must be executed in a specific order.
Intelligent Feedback Loop and Review Workflow: Mozzie implements a strict state machine (Draft -> Ready -> Running -> Review -> Done). During the "Review" phase, the tool provides a comprehensive diff viewer and terminal replay. If a user rejects a contribution, Mozzie injects the full attempt history and the specific rejection reason into the agent's next prompt. This "memory" mechanism ensures that agents learn from their mistakes within the context of a single task, preventing repetitive errors.
Local-First Architecture with LLM Orchestration: Unlike cloud-based platforms, Mozzie stores all data locally in a SQLite database via Rust (sqlx). It leverages a central "Orchestrator LLM" (OpenAI, Anthropic, or Gemini) to parse natural language instructions into actionable work items, which are then distributed to the specialized coding agents. This architecture ensures that sensitive code remains on the local machine while only the high-level orchestration requests are sent to the LLM provider.
Problems Solved
Context Switching and Sequential Latency: Developers often waste time waiting for a single AI agent to finish a task before starting the next. Mozzie eliminates this by allowing parallel task execution, effectively turning a single developer into a project manager.
Agent Collision and Merge Conflicts: When multiple AI tools or scripts edit the same file system, they often overwrite each other. Mozzie’s use of git worktrees ensures that every agent operates in a sandbox, with all changes merged only after human review.
Target Audience:
- Full-Stack Engineers: Who need to build front-end components and back-end logic simultaneously using AI.
- Open Source Maintainers: Who need to triage and address multiple issues across a repository in parallel.
- DevOps Specialists: Automating infrastructure-as-code updates across different environments.
- AI Researchers: Testing different coding models (Claude vs. Gemini) on the same task simultaneously to compare performance.
- Use Cases:
- Large-Scale Refactoring: Breaking a monolithic service into microservices by assigning different modules to different agents.
- Parallel Feature Development: Developing a UI component, its associated API endpoint, and its unit tests concurrently.
- Automated Bug Fixing: Feeding multiple bug reports into the orchestrator and having agents diagnose and fix them in separate worktrees.
Unique Advantages
Differentiation: Most AI coding tools are either IDE plugins (like Copilot) or standalone chat interfaces. Mozzie is a management platform that sits above these tools. It does not replace the agent; it coordinates them. Its local-first approach (SQLite, Rust, native Git) provides significantly higher performance and privacy compared to web-based AI orchestration platforms.
Key Innovation: The integration of "Stacked Branches" and "Sub-work-items." This allows for a hierarchical workflow where child tasks merge into parent tasks, which are then pushed to the origin as a single, cohesive Pull Request. This mirrors professional software engineering workflows more closely than any other AI agent tool currently available.
Frequently Asked Questions (FAQ)
Does Mozzie replace agents like Claude Code or Gemini CLI? No, Mozzie is an orchestrator. It is designed to work with your existing agents. It handles the "plumbing"—worktrees, task queuing, and context management—so you can run those agents more efficiently in parallel.
Where are my API keys and code data stored? Mozzie is local-first. Your code stays on your machine in standard Git worktrees. Metadata is stored in a local SQLite database. API keys are managed securely through the operating system's native keychain, ensuring that your credentials are never exposed in plaintext.
How does Mozzie handle agent errors or failures? Mozzie tracks the process lifecycle of every agent. If an agent fails or produces an error, the work item transitions back to a "Ready" state. You can then provide feedback, and Mozzie will re-inject that context into the next run, allowing the agent to self-correct based on your instructions.
