magi
Pass
Audited by Gen Agent Trust Hub on Jul 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes a custom runner script (magi-run.sh) to dispatch questions to external LLM provider CLIs in parallel.
- It employs a Perl-based wrapper to enforce wall-clock timeouts and manage process groups, ensuring that unresponsive models do not hang the main agent.
- Command-line arguments are constructed using element-wise substitution, avoiding shell expansion for file paths and prompt data to prevent injection attacks.
- [REMOTE_CODE_EXECUTION]: The documentation provides optional installation instructions for the Claude Code CLI using a remote script from Anthropic's official domain.
- [EXTERNAL_DOWNLOADS]: The skill provides setup instructions for official CLI tools from trusted vendors, including Anthropic, OpenAI, and xAI. These downloads are from well-known, legitimate services.
- [DATA_EXFILTRATION]: To protect sensitive data, the skill implements a 'Confidential' scope. For the Codex CLI, a wrapper script (codex-sage.sh) identifies and disables third-party MCP servers at runtime to prevent unauthorized data access by sub-agents.
- [SAFE]: The skill prioritizes auditability by maintaining a local run directory (defaulting to /tmp/magi-runs/) with owner-only permissions (umask 077). It stores all prompts, raw model outputs, and metadata, providing a clear audit trail for the council's decisions.
Audit Metadata