How SkillRelay works

For everyone (the 30-second version)

Your company has shared Claude skills — reusable instructions Claude follows for common workflows (filing bugs, writing reports, onboarding customers). You get them automatically. To change one, you don't file a ticket or edit a file: you tell Claude what's wrong, in a normal chat.

The only setup you personally ever do: add the SkillRelay connector to Claude once (Settings → Connectors → Add custom connector) using the URL on your team's Settings page, and sign in with your work email so your edits carry your name.

For the admin

The setup checklist walks you through it: create the workspace, let SkillRelay create a private GitHub repo in your org, install Anthropic's Claude GitHub App on that one repo, and point claude.ai's admin settings at it (add marketplace → Sync automatically → set the plugin to Required). Five minutes, once. Every future skill and edit ships with zero admin involvement, because the plugin already exists — only its contents change.

Approvals: off by default, there when you want them

By default, safety comes from undo, not gatekeeping: every change is an immutable version with a named author, anyone can revert anything in one sentence, and the full history is visible to everyone. A fix that waits three days on a reviewer is a fix nobody makes — so start here.

For teams that want a gate, an admin can turn on approvals in Settings. Then:

Security model, honestly stated

Beyond Claude

The library isn't locked to one vendor. Claude gets the best experience — skills auto-distribute through plugins — but the connector is a standard MCP server, so any MCP-capable tool (ChatGPT's developer-mode connectors, Cursor, and friends) can connect to the same URL and read or edit the same skills with the same per-person sign-in. For tools that only eat text, Settings has a one-click Markdown export of the whole library, ready to paste into project instructions or an AGENTS.md.

When something goes wrong

← Back to sign in