This document describes security considerations for the claudestuff plugin marketplace.

When you install a plugin from this marketplace, you're granting it the same permissions that Claude Code has in your environment:

Important: Plugins run with full Claude Code permissions. Only install plugins you trust.

Several plugins in this marketplace integrate with MCP (Model Context Protocol) servers:

• SimpleMem - Local memory storage
• OpenMemory - Persistent memory across sessions

1. Local vs Remote: All MCP servers in this marketplace are configured for local operation
2. Data Storage: MCP servers store data locally (typically in ~/.cache/ or configured directories)
3. No External Transmission: No data is sent to external services unless you explicitly configure it
4. Server Access: MCP servers run locally and are only accessible to Claude Code

• Do not store credentials, API keys, or passwords in files processed by plugins
• Do not process confidential documents without understanding data flow
• State files (like Claude-State-Tracking.md) may contain task summaries

Agents defined in plugins operate within Claude Code's permission system:

1. Tool Access: Agents can use any tool Claude Code has access to
2. No Privilege Escalation: Agents cannot access more than Claude Code itself
3. Audit Trail: All agent actions appear in the Claude Code session output
4. User Approval: Sensitive operations may prompt for confirmation depending on your Claude Code settings

If you discover a security vulnerability in this marketplace:

1. Do Not open a public GitHub issue
2. Email the maintainer directly (or use GitHub's private vulnerability reporting)
3. Include:
   • Description of the vulnerability
   • Steps to reproduce
   • Potential impact
   • Suggested fix (if you have one)

Yes, report these:

• Plugin code that exfiltrates data unexpectedly
• Command injection vulnerabilities
• Insecure default configurations
• Credential exposure in logs or state files
• MCP server misconfiguration allowing unauthorized access

Not security issues (but still bugs):

• Plugin fails to handle malformed input gracefully
• Error messages expose file paths
• Plugin requires more permissions than documented

This is a personal showcase, not a commercial product. Response times reflect that:

• Acknowledgment: Within 1 week
• Assessment: Within 2 weeks
• Fix (if applicable): Best effort, depends on severity

1. Review the source code - All plugins are open source; read what they do
2. Check the README - Understand what access the plugin needs
3. Test in isolation - Try plugins on non-sensitive data first

1. Monitor Claude Code output - Watch for unexpected file access or commands
2. Use minimal permissions - Configure Claude Code with appropriate sandboxing
3. Keep backups - Plugins can modify files; maintain vault backups

1. Local only - Don't expose MCP servers to network
2. Regular cleanup - Periodically review stored memories
3. Understand retention - Know how long data is kept