También en español: rosmarops.com/es/blog/rosmarops-0-3-0/
RosmarOps 0.3.0 is out. It doesn’t score any new agent yet, but it now sees four more: Codex, Antigravity, Hermes Agent and Claude Desktop. That changes the map more than it sounds.
An agent RosmarOps doesn’t evaluate can still keep keys on your disk. And another agent, one you do use every day, may be able to read them. That’s the part I cared about in this version.
What’s new
- It detects Codex, Antigravity, Hermes Agent and Claude Desktop. No score for them yet.
- It reads the keys they keep in plain files (Codex’s
auth.json, Hermes Agent’s.envfiles, and the MCP servers of Antigravity and Claude Desktop), and the report shows which of your other agents can read them. - It tells you when Claude Code or Cursor can rewrite another agent’s instructions, such as the
AGENTS.mdthat Codex reads.
Harder to fool
A program that reads your agents’ configuration has to be careful not to become the weak point itself. In this version, git never runs anything from the repositories it reads, PowerShell loads modules only from the Windows system folder, and every scheduled task is checked, not only the recent ones. If someone changes the weekly review task you set up, the report says so. And if a run fails, you get a red report, never silence; if it was the scheduled review, also a notification.
What it still doesn’t do
It doesn’t score the four new agents: it doesn’t read their approval or sandbox settings yet. It doesn’t read Claude Desktop installed from the Microsoft Store, which keeps its configuration in another folder. It still contains no networking code and sends no telemetry. It only reads.
Next
In the next version: where each MCP server sends data (the domain of remote servers, the program behind local ones), and a warning when a server starts pointing somewhere else.
Get it
Download it for Windows from GitHub. The download is attested and the build is reproducible: how to check it. If you used an earlier version, RosmarOps asks you to confirm the new one the first time.
I wrote this with AI assistants. The machine, the findings and the decision to publish are mine.