How OpsCrew’s AI works
A crew of AI agents does the DevOps work, on Claude (opens in a new tab) by Anthropic (opens in a new tab). A model is used only where judgment helps.
What it never does
These rules are part of the code, not a promise in a prompt. They hold whether you use OpsCrew in a terminal or in Claude Code (opens in a new tab).
It never changes anything without your approval
Every change is a plan you see first, with its cost and its SHA-256 hash. Your approval is bound to that hash, so a plan that changes needs a new approval.
It never sends a secret to the model
The AI steps send parts of your code, never a key or a password. Your secrets stay in your environment, in GitHub (opens in a new tab)’s encrypted secrets and in an encrypted recovery copy.
Claude cannot approve for you
In Claude Code, each plan waits in a dialog only you can answer, and Claude Code asks your permission before every tool that changes something, even in its auto modes.
It never deletes until you type the name
Removing a server or an app asks you to type its name first.
It never reports back to us
OpsCrew runs on your computer. There is no OpsCrew account and no telemetry: nothing is sent to OpsCrew.
Where Claude comes in
In Claude Code
Your own Claude Code session does the thinking and talks to you. OpsCrew’s plugin gives it safe tools that never call a model themselves, so it needs no extra key.
In a terminal
The Analyst and the Code-change agent use Claude with your own Anthropic API key or Claude subscription, or through Bedrock, Vertex AI or Foundry.
With no AI at all
Run analyze and adapt with --no-ai and no model is called. The security scan never uses one.
See it work on your own app.
Run npx opscrew init in your app's repository. It checks your tools and tokens first, for free.