The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/mcp_client.py script executes shell commands and arguments defined in configuration files (such as .mcp.json or mcp-config.json) using the stdio transport. This allows for the execution of arbitrary binaries and scripts provided in the configuration.
[EXTERNAL_DOWNLOADS]: The skill's documentation and reference files (e.g., references/mcp-servers.md) recommend using npx -y to run MCP servers. This command automatically downloads and executes packages from the npm registry at runtime.
[REMOTE_CODE_EXECUTION]: By design, the skill facilitates the execution of remote code through the stdio transport coupled with package runners like npx, uv, or docker as demonstrated in the provided configuration examples.
[CREDENTIALS_UNSAFE]: The skill is designed to handle and use highly sensitive credentials, including Zapier API keys, GitHub Personal Access Tokens, and database connection strings. It explicitly instructs the agent to request these secrets from the user and stores them in configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a generic interface to numerous external tools and services. If the agent processes untrusted data that includes instructions to use this skill, the agent could be influenced to call sensitive MCP tools with malicious arguments.
Ingestion points: Data ingested by the agent during conversation or tool use, which is then passed as arguments to the call command in scripts/mcp_client.py.
Boundary markers: The skill lacks explicit instructions for the agent to ignore or delimit embedded commands within the data being processed for MCP tool calls.
Capability inventory: Subprocess execution via stdio transport and network operations via sse and streamable_http transports.
Sanitization: Arguments are passed to the call_tool method; however, the content of these arguments is derived from the agent's context, which may contain unsanitized external content.

mcp-client