The Agent Skills Directory

[SAFE]: The skill implements a local Model Context Protocol (MCP) server for managing 2FA codes. The architecture is explicitly designed to keep sensitive authentication seeds on the local machine and out of the LLM's context, mitigating risks associated with conversation logging and summarization.
[COMMAND_EXECUTION]: The script scripts/totp_mcp_server.py uses subprocess.run with list-based arguments to interact with system keychains (macOS security and Linux secret-tool). This is a legitimate and safe method for retrieving stored secrets for the current user and does not involve shell injection vulnerabilities.
[DATA_EXFILTRATION]: While the skill accesses sensitive TOTP seeds, it does so only to generate the 6-digit codes required for logins. It explicitly excludes tools for exporting or reading the raw seeds, and documentation emphasizes that the tool should remain behind a manual approval gate in the agent client.
[INDIRECT_PROMPT_INJECTION]: The skill provides comprehensive documentation on the risks of indirect prompt injection. It acknowledges that an agent could be induced to call the get_totp_code tool by untrusted content (like a malicious webpage) and implements a mandatory allowlist (TOTP_MCP_ACCOUNTS) to restrict which accounts are accessible, even if the agent is compromised.
[EXTERNAL_DOWNLOADS]: The Python script uses PEP 723 inline metadata to pin the official mcp SDK (mcp>=2.1,<3). This ensures that the environment is consistent and uses verified versions of the protocol library.

totp-mcp-server