B

B — Some risk signals Graded October 16, 2026Minor concerns worth a quick check.Based on public adoption and transparency signals as of October 16, 2026. Our opinion, not a code audit or security guarantee.

AI agent skill · Developer tools, Communication

refly

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.

Bottom line: Signal-based grade (4/7 points): very widely starred on GitHub, public source code. This reflects adoption and transparency, not a code review.

Popularity score
44/100
Rank
#229 of 500
GitHub stars
7.5k

What it does

Refly is an open-source framework for building AI agent skills using visual workflow definitions that run across Claude Code, Cursor, Codex, and other AI coding environments. Skills are packaged as reusable, composable infrastructure components rather than prompt-based hacks, enabling deployment to Slack, Lark/Feishu, Lovable APIs, and custom integrations.

Who it’s for

AI builders and platform teams who want to standardize agent capabilities across multiple coding and chat environments without reinventing skills for each context. Executives seeking a maintainable, infrastructure-first alternative to prompt-based bot customization.

Copy-paste setup

Check these steps against the project’s own instructions before running anything.

Verify against the repo README at https://github.com/refly-ai/refly (opens external site) for current installation. Likely involves cloning the repository and following their quickstart; npm package metadata shows 0 weekly downloads, suggesting adoption is early or primarily GitHub-based. Check their documentation for Claude Desktop MCP server configuration if applicable.

Risks

  • Credentials scope: Skills may encode API keys or auth tokens; verify isolation boundaries and rotation policies.
  • Prompt injection: Workflows accepting dynamic user input could be exploited if inputs aren't properly sanitized before execution.
  • Supply-chain immaturity: Early-stage project (high star count but minimal npm adoption) means fewer security audits and rapid breaking changes.
  • Data exfiltration via skills: Reusable skills could silently log or exfiltrate data; code review each skill before production use.
  • Dependency risks: Pinned dependencies should be audited; unclear if the project maintains security patches for upstream Claude/Cursor integrations.

How to stay secure

Least-privilege permissions

  • Rotate API keys regularly – Change any passwords or access tokens that skills use to connect to external services (like Slack, databases) every 30–90 days, or immediately if a skill is deleted or team member leaves.
  • Limit what each skill can access – When creating a skill, grant it only the specific permissions it needs (e.g., "read messages" not "delete all messages"); audit existing skills quarterly to remove unused access.
  • Use separate credentials per environment – Keep different API keys for testing, staging, and production so a leaked test key doesn't compromise live data.
  • Enable multi-factor authentication (MFA) – Require MFA on any platform accounts that host or authenticate skills (GitHub, Claude Code, Slack workspace admin, etc.).

Network isolation

  • Restrict outbound connections – Configure firewall or network rules so skills can only communicate with approved external services (whitelist IP addresses or domains rather than allowing all outbound traffic).
  • Audit third-party integrations – Before connecting a skill to Slack, Lark, or an external API, verify the integration's privacy policy and confirm no unnecessary data flows outside your organization.
  • Use VPN or private networks for sensitive data – Route skill traffic through a corporate VPN or private network if skills access databases, internal tools, or customer information.

Monitoring

  • Log all skill executions – Record who ran which skill, when, and what data it accessed; review logs weekly for unusual patterns (skills running at odd times, accessing unexpected services).
  • Set alerts for permission changes – Notify your team immediately if a skill's access level is modified or a new skill is deployed; this catches unauthorized changes quickly.
  • Track dependency updates – Monitor when Refly, Claude, or underlying libraries release security patches and apply them within 2 weeks of release.

Red flags to watch for

  • Undocumented or unclear skills – If you cannot explain what a skill does, who built it, or what data it touches, treat it as high-risk and remove it until someone audits the code.
  • Skills that handle passwords or payment data – Any skill that stores, processes, or transmits passwords, credit card numbers, or health information should be rare and reviewed by security or compliance staff before use.
  • Rapid version changes or no maintenance – If Refly, a skill dependency, or a third-party integration releases multiple versions in short periods or stops releasing security updates, plan to migrate away.
  • Requests to loosen controls – Be skeptical of pressure to disable MFA, share credentials, or grant "admin to everything" access; this is how breaches happen.

Alternatives

  • Anthropic's built-in tool_use: Native Claude feature for function calling; simpler but less infrastructure-like than Refly.
  • LangChain agents: Mature Python framework with broad model support and tool orchestration; heavier than a skills builder.
  • Superflows: API-first agent framework with natural language to action; more closed-source and SaaS-oriented.
  • n8n/Make workflows: Visual automation platforms with bot-friendly outputs; less specialized for AI agent skills.

Sources