Monitor MCP (Model Context Protocol) Activity

The Model Context Protocol (MCP) is how AI assistants like Claude Desktop, Cursor, and VS Code connect to outside capabilities — file systems, databases, ticketing systems, internal APIs. Each connection is an MCP server the AI can call on its own. That autonomy is the risk: an AI agent with MCP access can read a file or a database and send the contents onward without a human ever seeing it happen.

With the desktop agent installed, ZeusLock watches MCP traffic on the endpoint, applies the same detection rules as everywhere else, and inventories every MCP server in use across your fleet.

What it does

• Scans both directions — what the AI sends to an MCP server, and (optionally) what the server sends back.
• Enforces your policy — a request that violates a Block rule is stopped and the AI client receives a clean protocol error it can handle gracefully; the user gets a desktop notification. On Alert, sensitive values are anonymized in place, exactly as in the browser.
• Fails secure — content ZeusLock cannot inspect is blocked rather than waved through.
• Records metadata only — each event stores the server, method, direction, decision, and a non-reversible content fingerprint. Raw content is never stored or displayed.

Remote MCP servers — automatic

MCP servers reached over HTTP are covered automatically wherever the desktop agent is running. The agent reads the MCP configuration of installed AI clients (Claude Desktop, Cursor, VS Code) about once a minute, recognizes the corresponding traffic, and inspects it — no configuration on your side.

Local MCP servers — wrap them with the shim

Some MCP servers run as local processes on the machine itself (over stdio) and never touch the network. To cover those, launch them through ZeusLock by editing the AI client's MCP configuration — the ZeusLock binary starts the real server and inspects everything that flows through:

{
  "mcpServers": {
    "local-files": {
      "command": "C:\\Program Files\\ZeusLock\\ZeusLock.exe",
      "args": ["mcp-shim", "--server-id", "local-files", "--",
               "node", "/path/to/server.js"]
    }
  }
}

The dashboard generates ready-to-paste snippets for each AI client and operating system — open Agents → MCP setup, pick the client (Claude Desktop, Cursor, VS Code) and the OS, and copy the block. The dashboard cannot rewrite configuration files on user machines — deploy the change with your IT tooling (Intune, Jamf, GPO scripts…), the same way you deploy the agent itself.

The MCP Monitoring page

MCP Monitoring in the sidebar gives you, over the last 7, 30, or 90 days:

• Headline counters — total MCP activity, blocked, alerts, active servers.
• Activity — every event, filterable by server, direction, decision, method, and user, with a detail view per event.
• Server inventory — every MCP server seen across the fleet, marked discovered (found in a client's configuration) or protected (traffic actually flowing through ZeusLock). Discovered-but-not-protected servers are your to-do list.
• Export — CSV or JSON, on Business and Enterprise plans.

Settings

MCP monitoring is active whenever the desktop agent is running. Two aspects are configurable per organization: whether server responses are scanned in addition to requests, and whether local stdio wrapping is enabled. Note that prompts inside MCP traffic are covered by the sensitive-data scan; the jailbreak classifier targets AI chat tools instead.