The Agent Skills Directory

[COMMAND_EXECUTION] (MEDIUM): The file scripts/connections.py implements MCPConnectionStdio, which utilizes the stdio_client from the mcp library to execute local processes.
Evidence: The _create_context method in MCPConnectionStdio passes command, args, and env parameters directly to StdioServerParameters for subprocess execution. This is a core feature of MCP but represents a significant security capability.
[EXTERNAL_DOWNLOADS] (LOW): The skill supports network-based transport layers for MCP servers.
Evidence: MCPConnectionSSE and MCPConnectionHTTP in scripts/connections.py allow connections to arbitrary external URLs. While standard for the protocol, this facilitates outbound communication.
[REMOTE_CODE_EXECUTION] (LOW): Indirect Prompt Injection surface identified.
Ingestion points: Untrusted data enters the agent context via the call_tool method in scripts/connections.py, which returns result.content from external MCP servers.
Boundary markers: Absent; the code does not implement delimiters or warnings for the agent to ignore instructions embedded in tool outputs.
Capability inventory: The skill can execute subprocesses (stdio_client) and perform network operations (sse_client, streamablehttp_client).
Sanitization: Absent; the response content from the MCP server is returned directly without validation or escaping.

mcp-builder