RemoteSkill plugin

imageGitHub Release

Points an agent at your RemoteSkill library: one MCP server at https://mcp.remoteskill.md/mcp, declared in both the Agent Plugins 1.0.0 format (plugin.json, mcp.json) and Claude Code's (.claude-plugin/, .mcp.json).

Sign-in is OAuth through your agent's own prompt. Skills are per-account content served behind that sign-in, so the plugin carries none; skills/remoteskill/ holds only the usage skill. A plugin installed before you sign in shows nothing until the sign-in completes, because the format treats an unauthorized server as a failed connection.

Tools

• list_skills: your catalog, every skill's name and full description, each with its ID and URL.
• read_skill: one skill's SKILL.md and file manifest, or a single bundled file.
• create_skill: a new hosted skill from the SKILL.md you give it, which carries the name in its frontmatter. An existing name is refused with every holder listed.
• write_skill_file, edit_skill_file, delete_skill_file: change one file in a hosted skill, by the skill's stable ID. Write sends the whole file; edit replaces one exact fragment.
• add_skills: add skills from links, each naming one skill: a GitHub link to its SKILL.md on a branch, or a RemoteSkill share link. A repository on its own is refused, because it may hold many skills. Results come back per link, in order, with any failure inline beside its link.
• remove_skills: remove whole skills by stable ID. Frees their names, ends every follower's access, and cannot be undone. An ID that matches nothing answers not_found, so a retry after a partial success is safe.

Writes reach exactly what you could edit in the app: hosted skills nothing else updates. A skill mirroring a GitHub repository or following someone's shared skill is read-only here.

Where it works

The server speaks MCP revision 2026-07-28 only. Each row carries the date it was measured, because they were not all measured on the same day:

Claude Code was measured on the wire: a local server logged its requests, and it sent MCP-Protocol-Version: 2026-07-28 with server/discover as its first call, never initialize. The claude.ai row is weaker evidence, measured in the product rather than on the wire: the connection is healthy and all seven tools are listed, split two and five by their read-only and destructive annotations, which needs a completed tools/list to render. Claude Desktop is not in the table because it was not tested.

The rows still at "not yet" connect without any plugin change once they ship the current revision, which is exactly what Claude Code did between 2.1.226 and 2.1.229.

Connecting

The server is OAuth, so a fresh install reports ! Needs authentication and shows no tools until you sign in. That is the first state, not a fault. SETUP.md walks an agent through the connection, and in Claude Code the manual path is /mcp, then plugin:remoteskill:remoteskill, then Authenticate.

Install

# Codex
codex plugin marketplace add LeeorNahum/RemoteSkill-Plugin
codex plugin add remoteskill@remoteskill

# Claude Code
/plugin marketplace add LeeorNahum/RemoteSkill-Plugin
/plugin install remoteskill@remoteskill

# Cursor
git clone https://github.com/LeeorNahum/RemoteSkill-Plugin ~/.cursor/plugins/local/remoteskill

VS Code: clone anywhere, then add the directory to chat.pluginLocations in user settings.

A test in the RemoteSkill repository asserts both manifests equal the endpoint the deployment advertises, so they cannot silently diverge.

Licensed under the Apache License 2.0.