RosmarOps

About RosmarOps

RosmarOps shows what the AI coding agents on your computer can do: what each one can read and change, which keys it can reach, where its MCP servers send data, and whether one agent can change another agent’s configuration or instructions. Everything is computed on your computer. It is made by Albert Garcia Pujadas.

What it is today

A free, open-source program (AGPL-3.0) for Windows and Linux, and for macOS in preview. You install it from the Microsoft Store, with Homebrew on macOS and Linux, or by downloading it from GitHub (download). It reads the configuration your agents leave on disk and gives you a one-sheet report: your computer’s control score, from 0 to 100, a class from A to D that says how much your agents do without asking you, and the findings that weigh most, with how much the score would go up if you change them.

If you don’t want to adjust permissions one by one, you tell it what you use your agents for and it suggests a profile for each one, with a text to paste into the agent itself. The agent shows you the change before saving it: RosmarOps touches nothing. If you want, you connect the program to your space at app.rosmarops.com and see the results of all your computers there.

It runs on Windows and Linux, and on macOS in preview. Today it evaluates Claude Code, Cursor, Hermes Agent, Codex, Grok Build, Antigravity CLI, GitHub Copilot in VS Code, Cline in VS Code and Gemini CLI. It detects Continue, Antigravity, Claude Desktop and n8n, but doesn’t score them yet. More on the home page.

What it never does

  • It sends nothing unless you approve it first. Without your permission it connects to nothing, not even to check for updates. If you connect it to your space, it sends only your results, encrypted, only to app.rosmarops.com.
  • It sends nothing unless you approve it. If something fails, it asks whether to send the list of errors: codes, version, system and day, never files, paths, names or keys.
  • It never changes your files, settings, keys or agents. The report tells you what to change; you make the change.
  • It never shows a key in full: only its first and last four characters.
  • It only writes in its own folder (~/.rosmarops/) and in the reports folder you choose.

You don’t have to take our word for it. The trust model lists every claim and how to check it. Each release comes with a GitHub attestation of the commit and workflow that built it, and building the same version from its source, with the Go version and command we document, gives the same program.

What it doesn’t do yet

  • Downloaded from GitHub, the program isn’t signed with a code-signing certificate yet: Windows may warn you the first time you open it. From the Microsoft Store it doesn’t. On macOS it isn’t notarised by Apple: with Homebrew nothing else is needed; if you download it from GitHub, the first time takes two commands in Terminal.
  • No independent security review has been published.
  • It reads configuration at rest. It doesn’t see what an agent does while it runs, settings kept only inside an app or on a vendor’s servers, or anything set only in one terminal session. What it couldn’t read is listed in the report and never counts as clean.
  • On macOS it is a preview: the report lists what it does not check yet on macOS, and none of that counts as clean.

What we publish

  • The Observatory: what each AI agent can do by default, company by company, what protections it has (sandbox, read-only mode, whether it asks first, network cut) and its security changes, each value with its date and source. With a published criterion, it says which agents to start with. No one pays to appear.
  • How to try an AI agent in a test environment: a guide to see what an agent could touch before you give it your folders.
  • The methodology: how the control score and the class are calculated, what they don’t measure and how we recommend an agent.
  • This blog: product news, in English, Spanish and Catalan.

Where it comes from

RosmarOps began in March 2025 as a research blog, in Spanish, around one question: what can an AI system do that you don’t expect, and who decides what it reads. It published more than a hundred articles along four lines:

  • Adversarial SEO and AI SEO: how what search engines and assistants read, cite and recommend gets manipulated.
  • Information warfare: how AI makes disinformation and influence operations cheaper.
  • AI security: attacks on models, data poisoning and prompt injection.
  • Agents and autonomy: what changes when AI stops answering and starts acting.

From March 2026 the blog focused more and more on agents, because with AI coding agents that question reached the computer of the people who build: an agent reads files, runs commands and edits code with your permissions. The product came out of that. To see how we got there (in Spanish):

Who is behind it

RosmarOps is made by Albert Garcia Pujadas, who has worked in digital, technology and marketing for more than 30 years and builds with AI agents every day. He writes at qtorb.com and is on LinkedIn.

How to read this blog

This blog publishes RosmarOps news, what your agent can do and the methodology, in English, Spanish and Catalan. The archive from before October 2026 was retired in November; the articles about agents that still hold are being rewritten here. The first post: I audited my own AI agents. My machine scored 18/100.