RosmarOps

What can Claude Code subagents do without asking you?

A Claude Code subagent can do whatever the header of its file allows: if the header lists no tools, it inherits every tool of the conversation, and it can have its own permission mode, its own commands and its own MCP servers.

What a subagent is

A subagent is a job with a name of its own. Claude Code hands it a task, such as reviewing code or searching the documentation, and the subagent does it with its own instructions. Each one is a Markdown file with a YAML header, in the project's .claude/agents/ folder or in yours, ~/.claude/agents/.

Four things the header decides

  • Its tools. If the header lists none, the subagent inherits every tool of the conversation: reading, writing, running commands, going online.
  • Its permission mode. A subagent can have its own, including modes that don't ask, such as acceptEdits, auto, dontAsk or bypassPermissions. There is a limit: bypassPermissions only works if the main conversation already has it, and if the main conversation is in bypassPermissions, acceptEdits or auto, the subagent takes that mode and its own is ignored.
  • Hooks. Commands that run on their own at certain moments, for example before or after a tool is used.
  • MCP servers. The header can define new servers, which connect as soon as the subagent starts.

Where the file lives matters

The hooks and MCP servers of a subagent in a project wait until you trust that folder. Those of a subagent in your home folder run in all your projects, without that question.

Why they get forgotten

A subagent is created once and forgotten. You copy it from a tutorial, a repository brings it or a tool installs it, and it stays there, with its permissions, every time Claude Code delegates.

How to use them with care

  1. Give each subagent only the tools it needs.
  2. Don't give it a mode that doesn't ask, unless you know why.
  3. Before you trust a project, open its agents folder and read the headers.
  4. Watch above all the ones in your home folder, because they go everywhere.
  5. Check now and then which subagents you have.

What RosmarOps tells you

Since version 0.17.0, RosmarOps reads only the header of each Claude Code subagent, never its instructions. It tells you when one works without asking you, runs a command of its own, connects to MCP servers defined in its file, or can use every tool. Subagents in your home folder run in every project without the folder-trust question, and the report says so.

Three of these four findings count in your score. A subagent that can use every tool is shown, but doesn't lower your score.

Read more

RosmarOps is free and open source. It only reads, and nothing leaves your computer unless you approve it.