OpsCrew in Claude Code: the plugin

The OpsCrew plugin lets you run OpsCrew from a Claude Code chat: the security scan, the interview, the design with real prices, the code changes and the deploy. Your Claude Code session does the reasoning and talks to you. OpsCrew’s MCP server supplies safe, deterministic tools and never calls a model itself, so the plugin needs no extra API key.

Every change still waits for your approval of its exact plan, in a dialog that only you can answer.

Install

Inside a Claude Code session:

/plugin marketplace add softzer0/opscrew
/plugin install opscrew@opscrew

Or from your shell:

claude plugin marketplace add softzer0/opscrew
claude plugin install opscrew@opscrew

The first command adds OpsCrew’s plugin marketplace, the softzer0/opscrew repository on GitHub. The second installs the opscrew plugin from it.

What it runs

When Claude Code installs the plugin, it also installs the OpsCrew CLI at the plugin’s version (opscrew from npm) and its dependencies, exactly as the plugin’s package-lock.json pins them, with npm and without running any install script. The plugin then starts OpsCrew’s MCP server on your computer from that copy:

node ${CLAUDE_PLUGIN_ROOT}/node_modules/opscrew/dist/opscrew.js mcp

You need Node.js 22 or newer and npm on your PATH. The download is about 37 MB. The MCP server never calls a model, so it needs none of the Claude Agent SDK that the CLI’s own AI steps install once, when you first run one.

The server runs with your shell’s environment, so it reads the same tokens as the CLI (HCLOUD_TOKEN, CLOUDFLARE_API_TOKEN, GITHUB_TOKEN and the others the guide’s pages name). Set them before you start Claude Code. Tool results never contain a token or a secret.

The skills

Skill	What it does
/opscrew:opscrew-deploy	The whole flow: scan, analyze, interview, design, prepare the code, provision and verify
/opscrew:opscrew-scan	The free security scan, with guided fixes
/opscrew:opscrew-adapt	Prepares the app’s code for its deployment
/opscrew:opscrew-troubleshoot	Finds out why the app is failing or slow, and what it uses: its status, figures, alerts and logs, and advice on what to change, read-only

Claude also uses them on its own when you ask things like “deploy my app”, “is my app secure?” or “why is my app failing?”. Every target works from the chat: a cloud server, a server you already have, your Portainer, your own computer, Cloudflare’s hosting and GitHub Pages.

Logs, figures and add-ons

• opscrew_status includes the resource figures and the scheduled check’s state and open alerts.
• opscrew_advise gives advise’s proposals on size, cost and upkeep, read-only. Applying one stays in your terminal: npx opscrew advise --apply <id>, then npx opscrew provision.
• opscrew_logs reads the app’s latest lines, read-only and masked: at most 500 lines and 64 KiB, and it never follows. For more, or --follow, run npx opscrew logs in your terminal.
• opscrew_addon_status, opscrew_addon_enable and opscrew_addon_disable show, turn on and turn off Beszel and Grafana Cloud, each plan approved by you in a dialog. Their tokens come only from OPSCREW_BESZEL_TOKEN and OPSCREW_GRAFANA_CLOUD_TOKEN in the MCP server’s environment, never from the chat, so Claude never sees them. Turning Beszel off from the chat keeps its history; deleting it, and opening Beszel’s dashboard, stay in your terminal (npx opscrew addon disable beszel, npx opscrew addon open beszel).

How approvals work

• Only you approve. A tool that changes something shows its plan, its hash and what it costs in a dialog (MCP elicitation). Claude cannot answer the dialog for you, and Claude Code asks your permission before every such tool call, even in its auto modes.
• Destroying plans need you to type the resource’s name in the dialog.
• A client without dialogs changes nothing: the tool stops and names the command to run in your terminal instead.
• Your app’s secrets never pass through the chat. You enter them in your own terminal (npx opscrew provision), or OpsCrew reads them from the file OPSCREW_APP_ENV_FILE names. Never paste a token or a secret into the chat.
• First connections: an SSH host key or a Portainer certificate is confirmed by you in a dialog, never by Claude.
• Removing an app (npx opscrew destroy) stays in your terminal.

Data and privacy

The MCP server runs on your computer, sends nothing to OpsCrew and has no telemetry. It talks only to the services your steps need, with your own tokens: see Data and the services it talks to in the plugin’s README, and OpsCrew’s privacy policy.

From the chat or from the terminal

Both work on the same files in your repository, so you can switch at any point: answer the interview in a terminal with npx opscrew interview (it keeps addresses you type out of the chat), then continue in Claude Code, or the other way round.

The plugin’s own page, with the files it ships, is the plugin’s README.