What is an AI agent orchestrator?
An AI agent orchestrator is a program that launches several AI agents at once, such as Claude Code or Codex, and hands each one part of the work.
What an orchestrator does
Paperclip, Orca, Paseo and Claude Squad are orchestrators. They organise agents like a team: Paperclip, for example, turns them into a company, with an org chart, tasks and a budget. Some give each agent a copy of your project on another branch, a git worktree, so their files stay apart.

Every agent it launches is still you
An agent launched by an orchestrator works as you. It sees your home folder and your keys, and it has the internet. The worktree keeps its files apart, not its reach.
Some launch them without asking you
- Paperclip starts Claude Code with
--dangerously-skip-permissions, so Claude Code asks you nothing. - Orca launches agents with full access by default.
- Claude Squad has an option to accept everything on its own. It is off by default.
- Paseo doesn't use the no-questions mode by default: it launches Claude Code in auto mode and Codex in auto mode with its sandbox.
An orchestrator is also a server
Paperclip listens on your computer. If someone makes it reachable from outside, anyone who reaches it can create an account, because sign-up is open by default. Versions before 2026.416.0 have known serious security flaws.
How to use one with care
- Don't leave the no-questions mode on by default: turn it on for a task and off afterwards.
- Turn on the sandbox in every agent it launches.
- Look at how it launches your agents.
- Make it listen only on your computer.
- Keep it up to date.
What RosmarOps tells you
Since version 0.17.0, RosmarOps finds Paperclip and tells you when it can put an agent such as Claude Code to work without asking you, when your version has a known serious security flaw, when it listens beyond this computer and when anyone who reaches it can create an account. It never opens Paperclip's database or its secret files: it knows them by their path.
For Paseo and Claude Squad, it reads from their own files the mode each agent is launched in. Orca keeps your choice in a database RosmarOps doesn't open, so the report says it assumes the factory setting, which gives agents full access.
Is a worktree a sandbox?
No. A worktree keeps each agent's files apart, but each agent still sees your home folder and your keys, and it has the internet. To keep their access apart, you need the sandbox in every agent the orchestrator launches.

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