Refly is an open-source platform that enables non-technical users and developers to build reusable, production-ready AI agent skills by describing tasks in natural language—without writing code. It solves the problem of fragile, one-off AI workflows by converting SOPs and manual processes into version-controlled, intervenable skills that can be shared, deployed, and monetized. Built with TypeScript and powered by a Model-Native DSL, Refly integrates with 3,000+ tools including Slack, Stripe, Salesforce, and GitHub, and exports skills to Claude Code, Cursor, Lovable APIs, and agent frameworks like LangChain and AutoGen.
Refly supports self-hosted deployment via Docker, offers a hosted workspace at refly.ai/workspace, and includes a community skill registry. Its architecture bridges visual no-code interfaces with enterprise-grade infrastructure, enabling teams to govern AI capabilities as shared assets rather than disposable prompts.
What You Get
- Vibe Workflow Builder - A copilot-driven visual editor that converts natural language descriptions like “summarize articles daily” into executable agent skills with no code.
- Deterministic Intervenable Runtime - Pause, audit, and re-steer AI workflows mid-execution to enforce business rules and recover from failures without restarting.
- Skill Registry & Versioning - Centralized, version-controlled repository for reusable agent skills, enabling teams to share, fork, and audit capabilities across projects.
- Export to Multiple Runtimes - Deploy skills as APIs for Lovable, webhooks for Slack/Lark/Feishu, or native tools for Claude Code and Cursor without reimplementation.
- 3,000+ Native Tool Integrations - Pre-built connectors for Stripe, Salesforce, GitHub, Notion, and more—no custom code needed to trigger actions or fetch data.
- Creator Rewards System - Publish your workflows and earn credits when others run them, turning personal automation into a monetizable asset.
Common Use Cases
- Building a viral content radar - A lifestyle creator uses Refly to automatically track trending posts, analyze viral structures, tag content, and save AI-generated topic ideas to Notion—replacing 2 hours of daily manual research.
- Automating e-commerce marketing - An e-commerce CEO deploys Refly to auto-generate personalized email campaigns, sync customer data from Shopify to CRM, and schedule social posts—boosting ROI by 60%.
- Creating a Slack bot for internal SOPs - A startup team builds a Refly-powered Slack bot that answers HR questions using company policy docs, reducing ticket volume by 40%.
- Exporting AI skills to Claude Code - A developer packages a data extraction skill built in Refly and deploys it as a native tool in Claude Code, enabling AI agents to pull real-time sales data without manual API calls.
Under The Hood
Architecture
- Monorepo structure enforces clear domain boundaries between backend (NestJS) and frontend (Next.js) via shared workspace packages, promoting modularity and reuse
- NestJS layers cleanly separate controllers, services, and data access with dependency injection, ensuring testability and loose coupling
- Cross-cutting concerns like authentication and rate limiting are encapsulated in reusable guards and services, avoiding duplication
- OpenTelemetry and Langfuse observability are abstracted behind clean interfaces, preserving separation of concerns
- Path aliases and consistent module resolution patterns reduce brittle imports and improve maintainability across packages
Tech Stack
- Backend powered by NestJS with TypeScript, leveraging BullMQ for job queues and Prisma for type-safe database operations
- Frontend built with Next.js and React, supported by Vitest and Cypress for comprehensive testing
- Data stack includes PostgreSQL, Redis for caching and real-time collaboration, and MinIO for object storage
- Observability integrates OpenTelemetry and Elasticsearch for tracing and logging
- Infrastructure is containerized via Docker Compose and enhanced with Gitpod for cloud-native development
Code Quality
- Extensive test coverage with Jest and Vitest, using type-aware mocks to validate behavior without external dependencies
- Strong TypeScript enforcement through Prisma-generated types and well-defined domain models, minimizing runtime errors
- Configuration-driven feature flags with type-safe parsing ensure robust environment-based control
- Consistent naming, modular structure, and explicit error handling patterns follow established best practices
- Linting and test patterns are standardized, with type-safe mocks preventing common runtime failures
What Makes It Unique
- AI skill packages are installed and activated as modular units with automatic dependency resolution, enabling dynamic workflow customization
- Distributed Redis-based locking ensures atomic skill installations across concurrent users, a rare capability in low-code AI platforms
- Extensible Markdown plugin system allows rich, interactive AI outputs rendered directly within markdown contexts using custom React components
- Canvas-integrated documentation modals provide in-context learning without disrupting workflow
- Unified schema-driven API layer enables seamless, type-safe communication between AI components and UI elements, reducing boilerplate