Wavefunction

Product Introduction

1. Definition: Wavefunction is a free, open-source desktop application for macOS that functions as a minimalist, parallel-processing coding agent. It is a local AI-powered development tool that leverages large language models (LLMs) to execute coding tasks autonomously within an isolated environment.
2. Core Value Proposition: Wavefunction exists to provide developers with a focused, non-intrusive AI coding assistant that prioritizes a clean workflow. Its core value is enabling parallel task execution within a single conversation while maintaining user control and privacy, contrasting with the trend of feature-bloated "super app" AI agents.

Main Features

1. Minimalist Desktop Interface: The app features a deliberately sparse UI focused on a central conversation panel and a simple project sidebar. It avoids built-in browsers, excessive tabs, and auxiliary panels, directing all external links to the user's default browser. This design philosophy reduces cognitive load and keeps the developer's primary focus on the code and dialogue.
2. Parallel Subagent Workflow: This is the core technical capability. When a user provides a complex task, Wavefunction can dynamically spawn multiple "subagents" to work on independent subtasks concurrently (e.g., API, UI, tests). Each subagent operates in its own Git worktree to prevent file collisions. The system can merge changes and, crucially, accept new instructions mid-task, allowing users to steer ongoing work without restarting it.
3. Isolated Computer Use via Docker: Instead of taking control of the user's actual desktop (mouse, keyboard, screen), Wavefunction performs "computer use" tasks—like web browsing or GUI interaction—inside a dedicated Docker container. This container provides a sandboxed desktop environment, ensuring the user's Mac remains untouched and secure. File transfers between the container and the host machine are explicit and user-approved.
4. Multi-Model & Local LLM Support: Wavefunction is model-agnostic. It supports signing into ChatGPT Plus/Pro, using API keys for providers like Anthropic, Google Gemini, OpenAI, OpenRouter, and Vercel AI Gateway, or running local models via Ollama, LM Studio, or any OpenAI-compatible server. Users can select and switch models per conversation.
5. Open-Source Prompt & Skill System: The project is MIT-licensed, with transparency extending "down to the prompt." It integrates with the Model Context Protocol (MCP) for tool connectivity and can automatically read project-specific instructions from AGENTS.md files and load skills from SKILL.md files in .agents directories, allowing deep customization and alignment with existing team workflows.

Problems Solved

1. Pain Point: Overly complex and intrusive AI coding assistants. Many AI developer tools become cluttered with features, tabs, and built-in tools that disrupt the developer's native environment and workflow.
2. Pain Point: Sequential, slow task execution. Traditional AI chat interfaces process tasks one at a time, forcing the user to wait for completion before providing the next instruction, which is inefficient for multi-faceted projects.
3. Target Audience: macOS developers and engineering teams who value a clean, focused workflow, demand privacy/security, and want to leverage AI assistance without ceding control of their local machine. It appeals particularly to developers using local LLMs or providers with zero-data-retention policies.
4. Use Cases: Rapidly prototyping a new feature with parallel development of backend, frontend, and tests; iterating on a complex refactor while providing continuous feedback; automating web-based research or data entry tasks without security risks; using a private, locally-hosted model for sensitive codebase interaction.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Unparalleled focus on a minimal, desktop-native UX; true parallel task execution within a chat; superior security model via Docker isolation for computer use; full model flexibility and support for local inference; completely free and open-source.
   • Cons: Currently macOS-only, limiting its user base; requires Docker Desktop for the computer use feature, adding setup complexity; the minimalist approach may lack integrated tools (e.g., built-in browser, terminal) that some users prefer for convenience.

2. Key Alternatives & Differentiation:

   • Cursor / Windsurf: These are full-fledged AI-powered IDEs. Wavefunction differentiates by being a standalone agent outside the IDE, adhering to a Unix-philosophy of doing one thing (agentic work) well, and not replacing the user's chosen editor.
   • OpenDevin / SmolAgent: These are prominent open-source coding agents. Wavefunction differentiates with its dedicated, polished macOS desktop application (vs. typically CLI/web-based), its patented parallel subagent architecture, and its containerized computer use model.
   • Claude Desktop / ChatGPT Desktop App: These are general-purpose AI chat applications. Wavefunction differentiates by being specialized for coding with parallel execution, project-aware skills, and a workflow designed for sustained development tasks rather than general Q&A.

Frequently Asked Questions (FAQ)

1. Is Wavefunction really free, and how does it work with paid models? Yes, Wavefunction is completely free and open-source under the MIT license. You pay only for the LLM services you use, such as your ChatGPT Plus subscription, API costs from OpenAI or Anthropic, or the computational resources for running local models like Llama 3 via Ollama.
2. How does Wavefunction's parallel coding work compared to a standard AI chat? Unlike standard sequential chats, Wavefunction can decompose a complex prompt (e.g., "build a dashboard with API and UI") and assign subtasks to multiple internal subagents that work simultaneously in separate Git worktrees. You can continue the conversation and issue new commands while these tasks run in the background.
3. What are the system requirements for running Wavefunction on Mac? You need a macOS machine capable of running the application. To use the "computer use" feature for web browsing and GUI automation, you must have Docker Desktop installed and running. For local model inference, you need sufficient RAM and CPU/GPU resources, typically via Ollama.
4. How does Wavefunction ensure the privacy and security of my code? Your conversations and settings are stored locally. Privacy depends on your model choice: using a local Ollama model keeps everything on-device. If using an API, you must rely on the provider's data policy; Wavefunction supports providers offering zero-data-retention (ZDR). The Docker container isolates all automated web activity from your host system.
5. Can I use Wavefunction with my existing development tools and scripts? Yes. Wavefunction can connect to external tools via the Model Context Protocol (MCP) and will read project-specific instructions from an AGENTS.md file in your repository. It also loads skill definitions from SKILL.md files in .agents folders, allowing you to customize its behavior for your specific codebase and workflows.