Gate the Access. Protect the Data.
AI agents reach enterprise apps through Model Context Protocol (MCP) servers, and an MCP gateway decides which apps they can access. Agents also move data through channels a gateway never sees, including shadow MCP servers, terminal commands, and agent-to-agent connections. This datasheet compares the MCP gateway and Cyberhaven’s endpoint sensor, and shows how they work together across an agentic workflow.
What's inside
Where MCP Gateways and Endpoint Sensors Differ
The gateway gates which apps agents reach through approved MCP servers. The endpoint sensor monitors and protects data on the endpoint, where agents and users work.
A gateway sees only the MCP servers registered with it. The endpoint sensor discovers registered and unregistered MCP servers on the device.
The gateway inspects MCP messages at a single point in time. The endpoint sensor inspects data continuously and captures its lineage.
Exfiltration Channels MCP Gateways Do Not Cover
Gateways monitor the traffic routed through them. Agents also reach unapproved AI models, shadow MCP servers, and agent-to-app connections through APIs, SDKs, and webhooks.
Agents run terminal commands, connect to other agents over the Agent2Agent Protocol (A2A), and submit data in browser sessions.
Users share agent output through email, web uploads, and USB, which moves data after the agent finishes its work.
How Both Layers Protect One Agentic Workflow
A numbered workflow diagram follows data from the employee prompt to agent output, and marks which control protects each of seven steps.
Three use cases cover discovering shadow MCP, allowing access while stopping data leaks, and completing the investigation.
A closing section for teams evaluating an MCP gateway shows what Cyberhaven adds to the agentic security stack, with a path to request a demo.