Refly Open-Source Agent Skills Builder: A New Paradigm for Cross-Platform AI Skill Development

Overview

Refly is an open-source Agent Skills Builder that has garnered over 7,200 stars on GitHub, making it one of the most watched emerging projects in the AI Agent development space. Its core philosophy treats "skills as infrastructure, not prompts" (Skills are infrastructure, not prompts), enabling developers to define skills through visual Vibe Workflows and run them on mainstream AI programming tools like Claude Code, Cursor, and Codex.

What Is an Agent Skills Builder?

From Prompts to Skills Infrastructure

Traditional AI Agent development often relies on carefully crafted prompts, but this approach has clear limitations: prompts are difficult to reuse, lack structured management, and have high migration costs across platforms.

To understand this pain point, we need to look back at the evolution of Prompt Engineering. Since ChatGPT ignited the large model application boom in 2022, prompt engineering has progressed through multiple stages—from simple instructions, Few-shot Learning, to Chain-of-Thought and the ReAct framework. However, even the most carefully designed prompts are fundamentally "unstructured natural language text"—they lack version control, interface constraints, and modular composition capabilities. When a team needs to maintain dozens or even hundreds of Agents, prompt management becomes a nightmare: minor wording changes can cause behavioral drift, capabilities cannot be shared between different Agents, and switching underlying models requires debugging each one individually.

Refly proposes an entirely new paradigm—abstracting Agent capabilities into "Skills" and building them as reusable, composable infrastructure. This approach isn't unfamiliar in software engineering: just as containerization evolved application deployment from manually configuring servers to standardized Docker images, Refly aims to elevate Agent capabilities from scattered prompt fragments into "skill units" with well-defined input/output interfaces that can be independently tested and version-managed.

This means developers no longer need to write prompts from scratch for every scenario. Instead, they can assemble existing skill modules like building blocks to quickly construct powerful AI Agents.

Vibe Workflow: Defining Workflows Intuitively

Refly introduces the concept of "Vibe Workflow" for defining skills. The term "Vibe" emphasizes an intuitive, low-barrier approach to workflow definition.

The concept of "Vibe Coding" was first coined by Andrej Karpathy—former OpenAI Research Director and Tesla AI lead—in early 2025. He described a programming style where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists"—developers simply describe their intent in natural language while AI handles code generation and debugging. This philosophy quickly resonated with the developer community, becoming one of the hottest tech trends of 2025. Refly's Vibe Workflow extends this philosophy to the realm of Agent skill orchestration.

From a technical implementation standpoint, visual workflow orchestration is not entirely new. Tools like LangGraph (a graph-structured Agent orchestration framework from the LangChain team), Dify (an open-source LLM application development platform), and Flowise have already provided similar visual capabilities. But Refly's differentiation lies in the fact that it's not just a workflow orchestration tool—it packages orchestration results into "skills" that can be directly invoked across multiple AI programming environments. This "define once, run everywhere" approach is closer to an infrastructure-level abstraction.

Developers can visually describe how an Agent should execute tasks without diving into underlying orchestration logic. This design philosophy aligns with the current "Vibe Coding" trend—lowering technical barriers so more people can participate in building AI applications.

Core Capabilities and Use Cases

Multi-Platform Compatible Execution

Skills built with Refly can run on multiple mainstream AI development platforms, including:

Claude Code: Anthropic's command-line AI programming tool that uses the terminal as its interaction interface, excelling at complex multi-file codebase operations. It runs directly in the developer's local environment, can read project context and execute shell commands, and is considered one of the most capable tools for "Agentic Coding" (autonomous programming).
Cursor: One of the most popular AI code editors currently available, deeply modified from VS Code with AI capabilities deeply integrated into every interaction point—Tab completion, inline editing, multi-file refactoring, and more—with a massive developer user base.
Codex: OpenAI's cloud-based code generation and execution platform (also known as Codex CLI) that can autonomously complete coding tasks in sandboxed environments, supports parallel processing of multiple tasks, and is suited for batch code generation and modification scenarios.

These three platforms represent three different forms of AI programming tools: command-line Agent, IDE integration, and cloud sandbox. To make the same set of skills work seamlessly across these three vastly different runtime environments, Refly needs sufficient abstraction at the skill definition layer to shield underlying platform differences. This cross-platform compatibility is a significant competitive advantage for Refly. Developers define a skill once and can use it seamlessly across different AI programming environments, avoiding platform lock-in.

Rich Application Scenarios

Based on the project description, Refly supports building various types of AI applications:

Clawdbot 🦞: An AI bot built on Refly's skill system
APIs for Lovable: Providing API interfaces for AI application development platforms like Lovable. Lovable is a popular AI full-stack application generation platform where users can generate complete Web applications through natural language descriptions. Refly providing API interfaces for it means the skill system can serve as backend capabilities called by higher-level AI application platforms.
Slack & Lark Bot: Intelligent bots for enterprise instant messaging platforms—these scenarios are in extremely high demand for enterprise AI deployment, covering high-frequency needs such as automated ticket processing, knowledge Q&A, and data queries.

These scenarios span a wide range of needs from individual development to enterprise collaboration, demonstrating Refly's versatility as skills infrastructure.

Technical Architecture and Ecosystem

TypeScript Tech Stack

Refly uses TypeScript as its primary development language—a wise technical choice. TypeScript has a massive developer base in the frontend and Node.js ecosystem, and its type system provides better structural support and developer experience for skill definitions.

Notably, TypeScript is becoming an increasingly mainstream choice in AI toolchains. Vercel's AI SDK, LangChain.js, and numerous MCP Server reference implementations all use TypeScript. There are several key reasons behind this: First, TypeScript's static type system is naturally suited for defining skill input/output Schemas, catching parameter type errors at compile time—particularly important for skill definitions that need to be passed between multiple platforms. Second, Node.js's asynchronous I/O model is well-suited for the heavy API calls and streaming response handling common in AI Agents. Finally, TypeScript's ecosystem overlaps significantly with Web development, making it easy to integrate skills into various Web applications and Bot platforms.

The project currently has 712 Forks, indicating strong community participation.

Open-Source Advantages

As the "first open-source Agent Skills Builder," Refly's open-source strategy is quite important:

Transparency: Developers can audit skill execution logic to ensure AI Agent behavior remains controllable
Extensibility: The community can contribute new skill modules, enriching the entire ecosystem
Self-hosting capability: Enterprises can deploy in private environments to meet data security requirements

Industry Trends and Outlook

Infrastructure-ification of Agent Development

Refly's emergence reflects an important trend in AI Agent development: evolving from manually writing prompts toward standardized, modular infrastructure.

Looking back at the evolution of AI Agent development paradigms helps us understand why this trend is inevitable. In early 2023, the ReAct (Reasoning + Acting) framework gave large models the ability to loop through "think-act-observe" cycles, but developers needed to manually orchestrate the logic of each step. Subsequently, OpenAI's Function Calling mechanism enabled models to call external tools in a structured way, but tool definition and management remained fragmented. In 2024, multi-Agent frameworks like AutoGPT and CrewAI emerged, making collaboration between Agents more complex and dramatically increasing the demand for standardized capability management. Refly was born precisely within this evolutionary trajectory—when the number and complexity of Agents reach a certain scale, "skills infrastructure" becomes an essential need.

As AI Agent use cases experience explosive growth, developers need more efficient tools to manage and reuse Agent capabilities.

The Possibility of a Skills Marketplace

Once skills are abstracted into standardized infrastructure, a natural extension is the emergence of a Skills Marketplace. Developers can publish, share, and trade skill modules they've built, forming an ecosystem similar to npm or Docker Hub. This would dramatically accelerate the development efficiency of AI Agent applications.

Relationship with the MCP Protocol

It's worth noting that Refly's skill system shares conceptual similarities with Anthropic's MCP (Model Context Protocol)—both attempt to provide standardized capability interfaces for AI Agents.

The MCP protocol was officially released by Anthropic in November 2024, defining a standardized communication protocol that enables AI models to connect to external data sources and tools in a unified manner. MCP uses a client-server architecture: AI applications (such as Claude Desktop, Cursor) serve as MCP clients, while various tools and data sources are packaged as MCP Servers. As of mid-2025, the MCP ecosystem has thousands of community-contributed Servers covering common scenarios like database queries, file operations, and API calls, with support from major AI vendors including OpenAI, Google, and Microsoft.

From an architectural perspective, MCP focuses more on "standardized protocols for tool invocation"—it solves the problem of "how an AI model calls a tool." Refly's skill system, on the other hand, focuses more on "standardization of workflow orchestration"—it solves the problem of "how multiple tools and steps combine into a complete capability." The two are not competitive but rather naturally complementary: Refly skills can call tools provided by MCP Servers, while the MCP protocol provides standardized interfaces for the underlying tool calls within Refly skills. How they co-evolve will be a direction worth continued observation.

Summary

Refly enters the AI Agent development space with the philosophy of "skills as infrastructure," lowering the barrier for skill definition through Vibe Workflow while achieving cross-platform compatible execution. In the current explosion of AI Agent applications, this approach to standardizing and modularizing Agent capabilities holds significant practical value. For developers building AI Agent applications, Refly is worth exploring in depth.

Key Takeaways

Refly is the first open-source Agent Skills Builder, treating skills as infrastructure rather than simple prompts, with over 7,200 GitHub stars
Implements intuitive skill definition through Vibe Workflow, lowering the barrier to AI Agent development
Supports cross-platform execution on multiple mainstream AI programming platforms including Claude Code, Cursor, and Codex
Covers various enterprise-grade application scenarios including API development and Slack/Lark bots
Reflects the industry trend of AI Agent development evolving from manual prompts toward standardized infrastructure