refly

Refly is an open-source platform for building what the project calls agent skills: structured, reusable pieces of workflow logic that AI agents can execute reliably. The core argument behind it is that most AI automation breaks down in practice because it relies on one-off prompts and fragile scripts. Refly is designed to let you define your business logic once in a visual interface and then deploy it anywhere that needs it.

The workflow starts in Refly's visual builder, where you describe what you want in plain language and the platform compiles it into a structured, version-controlled skill. Once built, a skill can be exported as an API endpoint (so an app builder like Lovable can call it), as a webhook for chat platforms like Slack or Lark, as a native tool for AI coding assistants like Claude Code or Cursor, or as a scheduled background job. The README describes this as skills being infrastructure rather than prompts.

Refly also has a concept of an intervenable runtime, meaning you can pause a running agent workflow mid-execution to audit what it is doing and redirect it if needed. This is aimed at enterprise use cases where an agent making an incorrect decision undetected could cause real problems. Version control and audit logs are built into the workspace for team use.

The platform connects to over 3,000 external tools out of the box including Stripe, Slack, Salesforce, and GitHub. It also supports the Model Context Protocol, which is a standard for connecting AI models to external data and tools. You can self-host Refly on your own server using Docker, or use the hosted version at refly.ai. A public skill registry at a companion repository allows the community to share, fork, and publish skills.

The project is written in TypeScript and released under the Refly AI License, which is a custom license rather than a standard open-source one. The full README is longer than what was shown.