The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation scripts (scripts/evaluation.py and scripts/connections.py) utilize the stdio transport mechanism to interact with MCP servers. This involves executing developer-defined CLI commands and arguments to launch local server processes for testing purposes.
[EXTERNAL_DOWNLOADS]: The documentation references documentation and SDK resources from the official Model Context Protocol organization on GitHub (github.com/modelcontextprotocol) and its primary domain (modelcontextprotocol.io). These are recognized and trusted sources within the development community.
[INDIRECT_PROMPT_INJECTION]: The testing harness ingests input from evaluation XML files and tool outputs from connected servers. This identifies a potential surface for indirect prompt injection where external data could attempt to influence the agent's logic, a consideration inherent to testing autonomous agents.
Ingestion points: Questions loaded from XML files in scripts/evaluation.py and result data returned by tool calls in the agent loop.
Boundary markers: None; the system prompt provides structural instructions using XML tags for the agent's output, but does not provide explicit delimiters for untrusted input.
Capability inventory: Execution of local server commands, invocation of tools on connected MCP servers, and network access to the Anthropic API.
Sanitization: The script processes question text and tool outputs without performing validation or escaping before passing them to the language model context.

mcp-builder