The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The SKILL.md file uses dynamic context injection via !command syntax to perform environment checks at load time. These commands check the local Python version and verify if the fastmcp library is installed. These are benign, informational operations for setup assistance.
[EXTERNAL_DOWNLOADS]: The scripts/requirements.txt file lists httpx2>=2.5.0 as a dependency. This is not a standard or officially recognized package from the httpx or Pydantic teams, despite claims in the documentation (references/migration.md). Requiring a non-existent or unverified version of a package poses a significant supply-chain risk.
[METADATA_POISONING]: The skill documentation describes a 'future' ecosystem (using dates in 2025 and 2026) and encourages the use of unreleased versions of libraries such as mcp>=2.0.0 and fastmcp>=4.0.0. This deceptive metadata could lead an agent or user to install untrusted software masquerading as official updates.
[DYNAMIC_EXECUTION]: The scripts/evaluation.py script facilitates the execution of tools from arbitrary MCP servers. While this is the intended functionality of an evaluation harness, it creates a vector for executing code provided by untrusted third-party servers.
[INDIRECT_PROMPT_INJECTION]: The evaluation harness in scripts/evaluation.py ingests data from external MCP tool outputs and XML-formatted evaluation pairs. Although it uses boundary markers (, , ), the lack of explicit sanitization for external tool outputs presents a surface for indirect prompt injection if the evaluated server returns malicious instructions.

fastmcp-creator