31INTEGRATIONS

MCP Security

MCP security can be an early wedge because it exposes the exact point where agents gain new capabilities.

Priority: P0
Phase: MVP
Primary owner: Platform / Security
Status: Blueprint

Objective

Model Context Protocol and similar tool protocols can dramatically increase agent capability by making tools discoverable and invokable. Agent Trust Cloud should act as a secure MCP gateway/registry that verifies servers, constrains tools and records every invocation.

What to build

Discover MCP servers and tool catalogs.
Verify server identity, ownership and environment.
Inspect tool schemas and classify side effects/sensitivity.
Proxy tool discovery and invocation through policy where possible.
Block unapproved servers, tools or parameter patterns.

Implementation decisions

Maintain trust status and last-seen/version fingerprints for each server.
Alert when a server changes tools or expands schema authority.
Apply egress/DLP checks to tool inputs and outputs.
Use allowlisted transports/endpoints and strong server authentication.

Definition of done

An unregistered MCP server cannot serve privileged production agents under enforcement mode.
Tool additions trigger inventory and policy review.
Every invocation maps to agent identity and server/tool version.

Success metrics

MCP servers discovered
Unapproved servers blocked
Tool-schema changes
MCP invocations governed

STRATEGIC FEATURE

MCP security can be an early wedge because it exposes the exact point where agents gain new capabilities.

This control is specified on page 31 of the blueprint.