Refly turns enterprise SOPs into versioned agent skills that leave as APIs
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.
At a glance
- What is it?
- Refly is a TypeScript monorepo where you describe business logic in natural language, a DSL compiles it into a versioned agent skill, and that skill leaves the building as an MCP server, an API, an SDK or a webhook. The friction worth knowing about sits in the engine requirements, the version numbers, and the export list.
- Who is it for?
- Worth a look if your team has written the same procedure into a script twice and wants a versioned artifact other agents can call. Check the Claude Code export path first if you need it, since the ecosystem list marks native Claude Code and Cursor export as coming soon while the routing table offers an Export for Claude Code guide, and skip it if you are still on an older Node release, because the manifest requires 24.0.0.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 13 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- refly-ai/refly7,556 ★ · +40
| Repository | 2026-09-13 | 2026-11-15 |
|---|---|---|
| refly-ai/refly | 7.5k | 7.6k |
40 Stars in the last 63 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Node 24 and pnpm 10 are the gate every install passes through
There is no single package to install. The workspace manifest declares the repo as refly-monorepo, marked private, at version 0.8.0, and sets two hard engine floors: pnpm >=10 and node >=24.0.0. Node 24 is the one that bites. A machine parked on an older long term support release cannot run the dev or build scripts at all. Because the repo is a pnpm workspace with pnpm-workspace.yaml over packages/ and apps/, npm install is not an alternative route either. Tasks run through turbo, and these are the script bodies the manifest defines:
turbo run dev
turbo run start
turbo run build
turbo run build --filter=@refly/api
turbo run build --filter=@refly/web
biome lint .Linting and formatting are biome, end to end tests go through cypress with a cy:open script to drive the runner, husky installs itself on prepare, and knip is wired up to find unused exports. Deployment is a separate path: a self deployment guide at docs.refly.ai walks through running Refly on your own server with Docker, and a hosted workspace sits at refly.ai/workspace if you would rather not.
The newest tag says v1.1.0 while the workspace manifest says 0.8.0
Version numbers disagree with each other here. The newest release is v1.1.0, published 2026-02-02, which comes after v0.10.0 from 2025-08-27 and v0.9.1 from 2025-08-14. The workspace manifest says 0.8.0. So a 1.x release line and a 0.8 manifest exist in the same repository, and the last push, dated 2026-11-02, lands after every tag, meaning main has moved about nine months past v1.1.0 with no release cutting it. The repository is not archived, so this is not a dead project, but the choice of a fixed point is awkward. Pin a tag and you take code from February. Track main and you take unreleased commits, which is what you get by default since the default branch is main. The license metadata is the same flavor of loose end: the repository record shows NOASSERTION as the license even though a LICENSE file sits at the top level.
Claude Code and Cursor export is the part marked coming soon
Not every destination on the list is finished. Under agent runtimes and platforms, native export for AI coding tools is listed as Claude Code and Cursor with the note coming soon. That is the route many readers care about most, since skills for Claude Code is one of the four headline use cases. The README is not consistent here, though: the Refly Skills section says skills can be shipped to Claude Code, and the routing table lists an Export for Claude Code path at 15 minutes. The destinations described as finished are stateful APIs that can power Lovable or a custom frontend, webhooks for Slack, Lark, Feishu and Microsoft Teams, and compatibility with agent frameworks including AutoGen, Manus, LangChain and custom Python stacks. Read the coming soon marker as the honest limit on this section. The wider list around it reads like a completed roadmap, and it is worth separating the two when you plan an integration.
One visual builder, then MCP servers, APIs, SDKs and webhooks
A skill is assembled in Refly's visual builder and then leaves in one of several shapes: deployed as an MCP server, exposed as a standard API, or shipped as a portable SDK. Workflows can also be run on a schedule with managed execution. On the input side the project counts 3,000+ native tools covering services such as Stripe, Slack, Salesforce and GitHub, and the full list of supported model and tool providers is kept in config/provider-catalog.json instead of prose, which is the file to check before you assume a connector exists. Model Context Protocol servers are supported directly, and private skill connectors are the route to your own databases, scripts and internal systems. The practical consequence for a reader is that breadth is not depth. A catalog entry confirms a connector, it does not tell you what parameters it takes, and the per API detail lives in the reference under docs/en/guide/api.
The intervenable runtime is the answer to the black box complaint
The runtime is where the complaint about opaque execution gets addressed. Skills execute inside a stateful runtime that can be paused, audited and re-steered while a run is still in progress, and business rules are described as enforced so that hallucinations are minimized and failure recovery is handled. Around that sits the governance layer: a central skill registry for securely managing, versioning and sharing capabilities, plus a team workspace with native version control and audit logs. A second repository, refly-skills, acts as the executable skill registry, where skills can be imported, forked and published by the community and run inside Refly with one click. The design puts review on something concrete, a versioned artifact rather than a prompt. It also means the guarantees a skill carries are exactly the guarantees of the rules someone wrote into it, so a thin rule set produces a thin guarantee.
The three minute build claim comes from the project, not from a measurement
Several numbers here are the vendor's own and deserve to be read that way. The summary says enterprise SOPs compile into executable agent skills, built in 3 minutes, and a capability bullet repeats a 3 minute deployment from a static SOP to a production ready skill. The efficiency claim stays qualitative: a streamlined DSL optimized for LLMs, ensuring fast execution and significantly lower token costs. No token figures or execution timings accompany those words in the capability list. The routing table is more concrete and more modest. It lists building a first workflow at 5 minutes, calling workflows via API at 10 minutes, connecting to Lark or Feishu at 15, exporting for Claude Code at 15, and building a Clawdbot at 20. Every one of those clocks starts after deployment, and deployment is its own Docker walkthrough.
n8n, Dify and LangChain each get a weakness assigned in the positioning table
The comparison section names competitors and gives each one a flaw. Workflow builders such as n8n and Dify are credited with good orchestration and then described as fragile, trigger only black boxes that are hard to reuse. Agent frameworks such as LangChain are credited with capable primitives and then described as demanding heavy engineering, manual boilerplate and constant maintenance. Refly's counter position is a streamlined DSL offering the speed of a graphical tool with the precision of code. Worth naming who wrote that comparison: it is the project's own, followed by a matrix of dimensions against legacy automation, so the table argues a position rather than measuring one. A team already running n8n or LangChain gets a claim to evaluate against its own experience, and the API reference plus the self deployment guide are where you find out whether the claim holds for your case.
Editorial conclusion
Worth a look if your team has written the same procedure into a script twice and wants a versioned artifact other agents can call. Check the Claude Code export path first if you need it, since the ecosystem list marks native Claude Code and Cursor export as coming soon while the routing table offers an Export for Claude Code guide, and skip it if you are still on an older Node release, because the manifest requires 24.0.0. Before committing, read the LICENSE file yourself, decide between tag v1.1.0 and main, and check the provider catalog for the one connector you actually need.
Frequently asked questions
what is refly
refly-ai/refly is an open-source platform for building versioned agent skills. It is a TypeScript pnpm monorepo named refly-monorepo at manifest version 0.8.0, requiring Node 24.0.0 or newer and pnpm 10 or newer, and skills are exported as MCP servers, standard APIs, portable SDKs, or webhooks for Slack, Lark, Feishu and Microsoft Teams.
is refly legit
The repository is public and still moving, with the last push dated 2026-11-02 and release tags v1.1.0, v0.10.0 and v0.9.1. One loose end is the license: the repository record shows NOASSERTION while a LICENSE file sits at the top level, so read that file before you pick a license for your own work.
is refly reliable
Refly states mechanisms rather than publishing measurements: an intervenable runtime that can be paused, audited and re-steered mid run, enforced business rules aimed at minimizing hallucinations and handling failure recovery, and native version control with audit logs in the team workspace. The lower token cost claim is stated without figures in the capability list.
is refly safe
Security is described rather than specified. The central skill registry is presented as securely managing, versioning and sharing capabilities, and private skill connectors reach your databases, scripts and internal systems. Self deployment puts Refly on your own server with Docker, so the exposure surface is yours to configure.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/refly-ai-refly)Star history chart for your README
[](https://hysenlabs.com/projects/refly-ai-refly)Was this useful?