The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches Model Context Protocol specifications and SDK documentation from official sources, including modelcontextprotocol.io and the protocol's official GitHub repository. These sources are considered trusted and well-known for this domain.
[COMMAND_EXECUTION]: The scripts/connections.py and scripts/evaluation.py scripts facilitate the execution of local shell commands to support the stdio transport mechanism. This allows the tool to launch and interact with MCP servers under development, which is a core feature of the protocol for local integration.
[DATA_EXFILTRATION]: The evaluation script handles sensitive configuration such as the ANTHROPIC_API_KEY and custom headers. These credentials are provided by the user through standard environment variables or command-line arguments and are used only for the intended communication with the model provider or the MCP server under test.
[PROMPT_INJECTION]: The scripts/evaluation.py harness processes outputs from external MCP servers, creating a surface for indirect prompt injection. . Ingestion points: External data enters through tool call responses in agent_loop. . Boundary markers: The EVALUATION_PROMPT enforces the use of XML tags (<summary>, <feedback>, <response>) to structure the model's output and distinguish it from external data. . Capability inventory: The script performs network calls to the Anthropic API and can execute local processes via stdio_client. . Sanitization: Content returned from tools is included in the conversation history without specific sanitization, relying on the structured XML output requirements to mitigate risks.

mcp-builder