Shadow AI is shadow IT for agents: developers installing and configuring MCP servers on their own machines, without review or approval. Any developer can add one from a GitHub repository, an npm package or a URL. There is no approval flow, no security scan and no record afterwards.
Danger
A misconfigured or malicious MCP server can send code, credentials or proprietary data back out through a tool response. The agent cannot tell the difference. It called a tool and got a result.
Every AI client keeps its MCP configuration in its own file:
| Client | Config |
|---|---|
| Claude Desktop | ~/Library/Application Support/Claude/claude_desktop_config.json |
| Claude Code | ~/.claude/settings.json |
| Cursor | ~/.cursor/mcp.json |
| VS Code | ~/.vscode/settings.json |
The mcphub CLI reads every known location and checks what it finds against the
servers your organization has approved.
Each discovered server is checked against the organization's registry, and
anything unrecognised is flagged with its name, transport, command and source.
mcphub report submits the result to the organization dashboard.
Most shadow-AI discovery reads an endpoint-security or device-management system you already run: CrowdStrike, Intune, Jamf. If you do not have one, and most teams under a few hundred people do not, those tools see nothing.
mcphub is developer-installed instead. The developer opts in by authenticating
with their organization, which means:
mcphub writes the MCP config in the first placeSubmitted reports aggregate into the dashboard: the total across all machines, the approved and unauthorized split, which unauthorized servers are spreading, per-developer status, and an alert when a new one appears.