The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a Python script (scripts/evaluation.py) designed to test MCP servers by executing them locally. It uses a connection utility (scripts/connections.py) that spawns subprocesses based on user-provided commands and arguments via the stdio transport. This is a standard requirement for local development and testing of MCP servers.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to integrate agents with external services via MCP tools. This architecture involves injecting output from these tools directly into the agent's context, which represents a potential surface for indirect prompt injection if the external data source is untrusted or compromised.
Ingestion points: Untrusted data enters the agent context through the call_tool method in scripts/evaluation.py and scripts/connections.py.
Boundary markers: The evaluation harness utilizes a system prompt that defines strict XML-like tags (<summary>, <feedback>, <response>) to help structure the agent's output, though it does not provide explicit isolation for tool content.
Capability inventory: The evaluation harness is capable of executing arbitrary local commands (to start servers), performing file operations (report generation), and making network requests to the Anthropic API.
Sanitization: Tool results are serialized to JSON strings before being passed to the agent; however, no advanced filtering or sanitization of potential instructions within that content is performed.
[EXTERNAL_DOWNLOADS]: The skill instructions direct the agent to fetch protocol documentation and SDK readmes from the official Model Context Protocol GitHub organization and their documentation website. These are treated as legitimate and neutral references for the technology being integrated.

mcp-builder