RosmarOps

What is an MCP server, and why does it give an AI agent power?

An MCP server gives an AI agent a new tool, such as reading your email or writing to your repository, and the agent can then do everything that server's key allows.

What MCP is

On its own, an AI agent can read and write text. To do things in the world it needs tools: reading your email, querying a database, opening a page, creating an issue. MCP is the standard protocol that connects an agent to tools and data from outside. Each MCP server is a new door you open for the agent.

Two kinds of door

A local server is a program the agent starts on your computer, with your user and everything you can do. A remote server is an internet address the agent connects to on your behalf.

The key

Almost every door has a key. To get into your email or your repository, the server needs a key for that service, and that key is often written in the same configuration file, next to the server, in plain text. The agent reaches exactly as far as the key does: with a server for your repository with write permission, it can write to your repository.

The less obvious part

Each server tells the agent which tools it has and what they are for. A third party writes that description, and the agent reads it as part of its instructions. A badly made or ill-intentioned server can slip orders in that way.

Who opens the doors

You, when you add a server. But also a project: a repository can bring its own file of servers. Claude Code loads servers from ~/.claude.json and from the repository's .mcp.json, and asks for your approval before using a project's servers. If you approve without looking, the door stays open.

How to use them with care

  1. Install servers only from people you know, and remove the ones you no longer use.
  2. Give each key the least permission: read only, if reading is enough.
  3. Don't accept a project's servers without seeing what they do.
  4. Keep keys out of the configuration file when the agent allows it.
  5. Check now and then which doors each agent has open.

What RosmarOps tells you

RosmarOps tells you which MCP servers each of your agents has, which file configured them and which credentials reach them, without showing the keys. Run it again and it tells you which servers are new and which now point somewhere else.

It reads the MCP configuration of Claude Code, Cursor, Cline, Codex, Grok Build, Gemini CLI, GitHub Copilot, Antigravity CLI and Hermes Agent.

Read more

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