The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill includes code templates for both TypeScript and Python that utilize the npx command to fetch and execute external code.
Evidence: The configuration snippets for mcpServers include the command npx claude-flow@alpha mcp start. This package is retrieved from the public NPM registry at runtime and is not associated with a known trusted organization.
[COMMAND_EXECUTION]: The skill instructs the creation of agents equipped with high-impact tools and recommends permission settings that remove human-in-the-loop oversight.
Evidence: The SDK implementation examples explicitly list Bash and Write in the allowedTools array.
Evidence: The agent-creator-process.dot file highlights a bypassPermissions mode described as "Auto-approve all," which poses a significant risk of unauthorized command execution if the agent is compromised or follows a malicious prompt.
[REMOTE_CODE_EXECUTION]: The framework encourages a pattern of "reverse engineering" and executing remote scripts as part of its "Phase 4: Deep Technical Enhancement."
Evidence: The integration of the claude-flow MCP server via npx constitutes remote code execution, as the logic of the server is hosted externally and executed locally to provide the agent's core capabilities.

agent-creator