The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to manage and execute local MCP servers as subprocesses through the StdioClientTransport. It executes commands and arguments specified in the .claude/.mcp.json configuration file, which is the standard mechanism for local MCP integration.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions to install several well-known and reputable tools, such as the gemini-cli and official MCP server packages (e.g., @modelcontextprotocol/server-memory, @modelcontextprotocol/server-filesystem, @modelcontextprotocol/server-brave-search) using standard package managers like npm and npx.
[PROMPT_INJECTION]: The skill ingests tool schemas, prompt templates, and resource content from external MCP servers, which are then integrated into the agent's context. This creates a potential surface for indirect prompt injection from server-provided metadata.
Ingestion points: Functionality for listing tools, prompts, and resources in scripts/mcp-client.ts (lines 92, 115, 137) and the static catalog in assets/tools.json.
Boundary markers: No explicit delimiter or instruction-guarding markers are implemented in the provided scripts to wrap content retrieved from external servers.
Capability inventory: Subprocess execution via StdioClientTransport in scripts/mcp-client.ts (line 59) and local file system access in scripts/mcp-client.ts (line 46).
Sanitization: Tool descriptions and resource contents are passed directly into the agent context without additional filtering or escaping.

mcp-management