Today's GitHub Trending tells a clear story: Agent infrastructure is splitting into layers.
anomalyco/opencode is fighting for the coding agent entry point. stablyai/orca manages a fleet of agents across desktop, mobile, and remote runtimes. akitaonrails/ai-memory provides long-term memory and handoff between different agent vendors.
Each solves a real pain point. But together they reveal a missing layer.
The missing layer: skill registry
When your agent fleet calls OCR, translation, voice transcription, or proofreading skills — who tracks which version of each skill is running? Who controls which agent has permission to call which skill? Who audits the calls?
The answer today, for most teams, is: nobody. Skills are scattered across repos, versions are untracked, permissions are wide open, and there's no audit trail.
What a skill registry actually does
This is where iflytek/skillhub comes in. It's a self-hosted, open-source skill registry for AI agents that provides:
- Skill package publishing + version management: Every skill has a version. You can roll back. You can pin.
- RBAC permissions: Not every agent should call every skill. Role-based access control keeps things locked down.
- Audit logs: Every skill call is logged. You know who called what, when, and which version.
How it fits with the rest of the stack
Think of the agent infrastructure stack like this:
| Layer | What it manages | Example repo |
|---|---|---|
| Agent entry point | Who runs the agent | anomalyco/opencode |
| Agent memory | What agents remember across sessions | akitaonrails/ai-memory |
| Agent fleet | How multiple agents are managed across devices | stablyai/orca |
| Skill registry | What agents can do, who allows it, and what version | iflytek/skillhub |
| Workflow orchestration | How skills are chained into workflows | iflytek/astron-agent |
The skill registry layer is the one most teams don't realize they need — until they have three agents calling five different skill versions and nobody knows which one broke production.
Getting started
If you're already running multiple agents and managing skills manually, skillhub is designed to be self-hosted and production-ready:
- Repo: https://github.com/iflytek/skillhub
- For workflow orchestration on top of registered skills: https://github.com/iflytek/astron-agent
When your agents become a fleet, memory and fleet management aren't enough. You need to know what they can do, who allowed it, and which version they're running.
Top comments (0)