Plugins, skills, and hooks for Claude Code. No customer data. No credentials. No secrets. Skills describe workflows and call MCP-connected tools (Salesforce, Slack, Gong, etc.) at runtime using credentials that live on the installing user's machine.

• Ship credentials, tokens, or API keys
• Store customer data
• Phone home to any external service
• Execute network calls on its own — only the MCP servers the user has authorized

When a portfolio company installs a plugin from this marketplace:

1. Code runs locally. Plugin code executes inside Claude Code on the user's machine. No cloud middleware.
2. Tool access is user-scoped. Plugins only access tools the user has connected via MCP / OAuth on their own machine.
3. Hooks are auditable. Every hook is a shell script in <plugin>/hooks/*.sh. Read them before installing.
4. Cache model. Plugins are copied into ~/.claude/plugins/cache/ on install — what you install is what runs. Edits to source require an explicit /plugin update to take effect.
5. No auto-update unless enabled. Auto-update can be toggled per marketplace in /plugin settings.

If you find a security issue:

1. Do NOT open a public GitHub issue.
2. Email the maintainer privately (contact via Radian Capital).
3. Include: affected plugin/skill, reproduction steps, and severity assessment.

You should expect an acknowledgment within 5 business days.

• Hooks that escape their intended scope (e.g., reading files outside the plugin folder when they shouldn't)
• Skills that leak credentials in their outputs
• Hooks that fail-open in dangerous ways (e.g., verify-connectors returning continue when it should block)
• Any path that writes data to a third party not explicitly authorized by the user

• Skills that produce wrong recommendations (file as a bug, not a security issue)
• Workflows that depend on connectors the user hasn't installed
• Performance issues
• Compatibility with non-Claude-Code harnesses