The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides an execute_mcp_code tool designed to run JavaScript and TypeScript code generated by the agent. The implementation relies on the Node.js vm module for sandboxing, which the accompanying security documentation explicitly warns is insecure and vulnerable to prototype pollution and constructor-based escapes. These vulnerabilities allow executed code to break out of the sandbox and access the host's process, require, and filesystem objects.
[COMMAND_EXECUTION]: The orchestrator's primary function is to spawn child processes using commands and arguments retrieved from a local mcp.registry.json file. This architecture allows for the execution of arbitrary system commands. If the registry file is manipulated or contains malicious entries, it can be used to execute unauthorized programs with the privileges of the agent process.
[EXTERNAL_DOWNLOADS]: The skill's documentation and registry schema support the use of npx and uvx to dynamically download and execute packages from the npm and PyPI registries. This introduces a supply chain risk, as the skill may fetch and run unverified or malicious remote code at runtime without explicit integrity checks or pinning.
[DATA_EXFILTRATION]: The skill implements an HTTP transport layer that sends JSON-RPC requests to external URLs defined in the registry. The documentation notes that these requests may include sensitive information such as API keys in the Authorization headers. If the configured URL is malicious or a configured MCP server is compromised, this mechanism serves as a direct channel for credential and data exfiltration.

mcp-dynamic-orchestrator