Find and Govern the AI Tools on Your Workstations
See every AI agent, skill, MCP server, model, and repository your team's workstations use, and allow or deny each one to catch shadow AI.
Developers install AI agents like Claude Code, Cursor, and GitHub Copilot, then extend them with skills and MCP servers that can read code, run commands, and reach external services. AI Security shows you all of it in one place: which agents run on which machines, what each one can use, and which repositories it works in. You review each agent, skill, and MCP server, and allow or deny it, so you catch shadow AI before it touches your code and data.
How Mondoo finds AI tools
Mondoo collects the AI inventory every time it scans a workstation. You don't need to enable a policy or a query pack: any macOS, Windows, or Linux workstation that you scan with cnspec reports its AI tools. To start scanning workstations, read Secure Operating Systems.
Mondoo finds:
| Component | What Mondoo collects |
|---|---|
| Agents | AI coding agents and assistants, such as Claude Code, Claude desktop, Cursor, GitHub Copilot, Windsurf, Cline, Gemini, and Aider, with their versions. |
| Skills | The skills installed for each agent, with their versions. |
| MCP servers | The MCP servers each agent is configured to use. A server set up in many projects is listed once. |
| Models | The models each agent has used, with token counts and cost. |
| Repos | The repositories each agent works in. Several checkouts of the same repository are listed once. |
Mondoo updates the inventory with each scan. When a developer removes a skill or an MCP server, it disappears from AI Security after the next scan of that workstation.
AI Security uploads the names of the skills, MCP servers, and repositories it finds, and the account each agent is signed in with.
Open AI Security
-
In the Mondoo App, navigate to the space you want to look at. To see every space at once, stay at the organization level.
-
In the left navigation, open the view switcher (it shows All technologies by default) and select AI Security.
AI Security is on for every organization. If it isn't in the list, contact Mondoo support.
-
The Dashboard opens. It shows:
- AI components: how many agents, skills, MCP servers, and models Mondoo found.
- Assets running agents: how many workstations run at least one AI agent.
- Denied: how many agents, skills, and MCP servers you denied.
- Awaiting review: how many agents, skills, and MCP servers have no decision yet.
- Top agents, Top skills, and Top MCP servers: the most widely used of each.
Explore what your team uses
The left navigation has a page for each kind of component. On each page, use Group by to change how the list is organized, and select a row to see its details.
| Page | What it shows |
|---|---|
| Agents | Each agent with its version, how many MCP servers and skills it uses, how many assets run it, its risk, and its status. Group by Version, Vendor, or Origin. An agent's page shows its version distribution and every installation. |
| Skills | Each skill with its versions, the agents that use it, the assets it's on, its risk, and its status. |
| MCP servers | Each MCP server with its source, the assets that use it, its tools, prompts, and resources, and its status. |
| Models | Each model with its provider, the agents that use it, tokens, and cost. The summary shows the cumulative model spend. |
| Repos | Each repository an agent works in, with the agents and assets involved. |
| Graph | How each agent connects to its skills, MCP servers, models, and repositories. Color shows risk and the outline shows status. Use Filter to show only some risks, statuses, or node types. |
Allow or deny an agent, skill, or MCP server
Every agent, skill, and MCP server starts as Unclassified. Review each one and record a decision.
-
Open Agents, Skills, or MCP servers.
-
In the Status column, select the item's status.
-
Choose a decision:
- Allow: the item is approved for use.
- Deny agent, Deny skill, or Deny server: the item isn't approved. Denied items count toward Denied on the dashboard.
- Mark out of scope: the item isn't something you govern.
Mondoo confirms the decision and updates the status right away.
To undo a decision, select the status again and choose Clear approval. The item goes back to Unclassified.
Where decisions apply
- A decision you make in a space applies to that space only. The same item still shows as Unclassified at the organization level and in other spaces.
- A decision you make at the organization level applies to every space in the organization, and overrides a decision made in a space.
- When an item leaves the scope, for example because the developer removed a denied skill from their workstation, its decision goes with it. If the item comes back, it starts as Unclassified again.
Models and repositories don't take decisions. To control them, decide on the agents that use them.
Related
- Connect Claude and other AI tools to Mondoo with the MCP server: ask an AI assistant about your Mondoo findings.
- Inventory your assets: browse every asset and the software installed on it.