refly
refly-ai · refly-ai/refly
An open-source platform for building, versioning, and deploying custom agent skills to various AI coding agents.
The first open-source agent skills builder. Define skills by vibe workflow, run on Claude Code, Cursor, Codex & more. Build Clawdbot 🦞· APIs for Lovable · Bots for Slack & Lark/Feishu · Skills are infrastructure, not prompts.
Preview
What you can do with it
- Create a custom skill for your Claude Code agent to automate a specific business process.
- Deploy a Refly-built skill as an API to integrate with external applications.
- Share and reuse versioned agent skills across different AI agent environments or teams.
Safety check
Usable, but read the notes first.
- ●No clear license: check with the author before using it in a commercial product.
Automated scan of the repo files. Always review code you run.
From the README
Refly — Agent Skills Builder Powered by Vibe Workflow
English · 中文
APIs for Lovable · Webhooks for Lark/Feishu · Skills for Claude Code · Build Clawdbot
Skills are not prompts. They are durable infrastructure.
Refly is the first open-source platform for building stable, atomic, and versioned agent skills.
Try it Now
Refly Skills
Refly Skills is the official executable skill registry for Refly.
- ⚡ Run instantly: Execute skills in Refly with one click
- 🧩 Reusable infrastructure: Versioned skills, not one-off prompts
- 🔌 Export anywhere: Ship skills to Claude Code or deploy as APIs
- 🌍 Community-powered: Import, fork, and publish your own skills
Explore the registry: Refly Skills Repo
Skills are deterministic agent capabilities—reusable across workflows, teams, and runtimes.
TL;DR: Refly compiles your enterprise SOPs into executable agent skills. Built in 3 minutes. Shipped anywhere.
Quick Start
Deploy Refly
-
📘 Self-Deployment Guide
(Recommended for Developers) Step-by-step guide to deploying Refly on your own server using Docker. -
🔌 API Reference
Complete API documentation for integrating Refly into your applications.
[!TIP] Want to explore instantly? Open the hosted workspace: Refly Workspace
What's Next?
After deployment, choose your path based on your use case:
| I want to... | Start here | Time |
|---|---|---|
| 🔧 Build my first workflow | Create a Workflow | 5 mins |
| 🔌 Call workflows via API | API Integration | 10 mins |
| 💬 Connect to Lark/Feishu | Webhook Setup | 15 mins |
| 🤖 Export for Claude Code | Export Skills | 15 mins |
| 🦞 Build a Clawdbot | Build Clawdbot | 20 mins |
Why Refly?
Most AI Agents fail in production because they rely on "Vibe-coded" scripts and fragile, black-box logic. As the ecosystem moves toward agentic frameworks like Claude Code, AutoGen, and MCP, the bottleneck is no longer the LLM—it's the lack of standardized, reliable actions.
Refly bridges the gap between raw APIs and intelligent agents. We allow you to codify messy business logic into structured, version-controlled Agent skills that any agent can invoke with 100% reliability.
Stop hard-coding tools. Build modular skills once in Refly's visual IDE and deploy them as MCP servers, standard APIs, or portable SDKs to any agent framework. Try it Now
Core Capabilities
🎯 Construct with Vibe (Copilot-led Builder)
Describe your business logic in natural language, and Refly's Model-Native DSL compiles your intent into a high-performance skill.
- Intent-Driven Construction: Describe the work once; Refly turns intent into deterministic, reusable, and composable skills.
- Efficiency at Scale: Streamlined DSL optimized for LLMs, ensuring fast execution and significantly lower token costs.
- 3-Minute Deployment: Transition from a static enterprise SOP to a production-ready agent skill in under 3 minutes.
⚡ Execute with Control (Intervenable Runtime)
Break the "black box" of AI execution with a stateful runtime designed for deterministic reliability.
- Intervenable Runtime: Pause, audit, and re-steer agent logic mid-run to ensure 100% operational compliance.
- Deterministic Guarantees: Enforce strict business rules that minimize hallucinations and handle failure recovery.
🚀 Ship to Production (Unified Agent Stack)
Unify MCP integrations, tools, models, and reusable skills into a single execution layer.
- Universal Delivery: Export as APIs for Lovable, webhooks for Slack or Lark/Feishu, or native tools for Claude Code and Cursor.
- Stable Scheduling: Run workflows reliably on schedule with managed execution.
🏛️ Govern as Assets (Skill Registry)
Transform fragile scripts into governed, shared infrastructure across your organization.
- Central Skill Registry: Securely manage, version, and share agent capabilities.
- Team Workspace Collaboration: Build together with native version control and audit logs.
Ecosystem
Refly is designed to be the universal bridge between your existing enterprise toolchain and the next generation of agentic runtimes.
Tooling & Protocols (Inputs)
Bring your own data and logic into Refly with zero friction.
- 3,000+ Native Tools: Seamlessly integrate with Stripe, Slack, Salesforce, GitHub, etc.
A full list of supported model and tools providers can be found in provider-catalog.json.
- MCP Support: Full compatibility with Model Context Protocol servers
- Private Skill Connectors: Connect to your databases, scripts, and internal systems
Agent Runtimes & Platforms (Outputs)
Export your deterministic skills to any environment where work happens.
- AI Coding Tools: Native export for Claude Code and Cursor (coming soon)
- App Builders: Power Lovable or custom frontends via stateful APIs
- Automation Hubs: Deploy as webhooks for Slack, Lark/Feishu, or Microsoft Teams
- Agent Frameworks: Compatible with AutoGen, Manus, LangChain, and custom Python stacks
Why Teams Choose Refly
For Builders: From Vibe to Production
Most agent tooling today falls into two categories:
- Workflow builders (n8n, Dify): Great for orchestration, but workflows are fragile, trigger-only "black boxes," and hard to reuse.
- Agent fram