The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's installation guide recommends piping a remote script from an untrusted repository into the shell (curl -fsSL https://raw.githubusercontent.com/xyTom/coding-tools-mcp/main/scripts/install.sh | bash), allowing for arbitrary code execution without verification.
[COMMAND_EXECUTION]: The skill provides an exec_command tool allowing the agent to run shell commands, with specific modes that bypass safety gates. The Dangerous mode explicitly disables all permission gates for command execution.
[EXTERNAL_DOWNLOADS]: The skill includes functionality and scripts to download third-party tunneling binaries (cloudflared, ngrok, and devtunnel) from remote sources to facilitate workspace access.
[METADATA_POISONING]: The skill content attributes the work to 'ara.so', which conflicts with the provided author context ('reason-machines'), potentially misleading users about the skill's origin.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted local files while possessing powerful modification and execution tools.
Ingestion points: Tools like read_file, list_files, search_text, and git_log read workspace content.
Boundary markers: None identified for content separation.
Capability inventory: Includes high-impact tools such as exec_command and apply_patch.
Sanitization: Path validation and basic environment filtering exist, but workspace content is not sanitized against embedded instructions.

coding-tools-mcp-server