Claude Mods Run Unsandboxed with User Permissions, Dan McAteer Reports
Summary
Claude Mods are plugins that run code inside Claude Code and change how the harness shows an agent's work to its operator, Dan McAteer reports in Latent Space. Anthropic ships six built-in mods, including you-should-know, an off-by-default side agent that flags missed details. McAteer says mods run unsandboxed with user permissions, and one check costs about 5,000 to 6,000 tokens.
Key Points
- Mods are TypeScript or JavaScript functions that respond to prompts, tool calls, and interface renders inside Claude Code.
- McAteer reports that 99 of his 111 Claude Code turns since October 15 met Next Steps Supervisor's threshold for a check.
- A built-in plugin-authoring skill teaches Claude Code to write mods from a plain-language request.