Security and Validation

The MCP Registry uses a governance-based security model to ensure that only trusted MCP servers can access your systems. This applies to both Public MCPs and Custom MCPs registered from your Artifactory.

How Security Works

Access control happens at two levels:

MCP-level: Which MCP servers are allowed in your project (governance).
Tool-level: Which specific tools within an MCP can be used (tool policies).

❗️
Important
MCP servers execute with the permissions of the user who invokes them through the Agent Guard. This means:

Users can only access data and perform actions they are already authorized to do.

MCP servers cannot escalate privileges beyond the invoking user's permissions.

All actions are auditable and traceable to specific users.

Validation at Registration

Local custom MCP servers are validated before they are added to the Registry: the package must exist in a local Artifactory repository, the repository must belong to the target project, and you must have ALLOW permission on the project and READ access to the repository. Remote custom MCP registrations enforce endpoint and header schema constraints (for example, unique header names). Public MCPs are scanned (e.g. for vulnerabilities and licenses) before you allow them for a project.

Remote MCP Header Schema

For remote custom MCP servers, the Agent Guard acts as a secure proxy to your external HTTPS endpoint. Administrators define which HTTP headers the proxy must send and whether each header is required or treated as a secret. Registration stores only that schema (header names and rules), not header values.

When a developer connects an agent, the Agent Guard prompts for any required header values, masks inputs marked as secret, and attaches the headers to every proxied request. Header values are not published in the AI Catalog. This keeps credentials out of registry configuration and ties header use to the connecting user's governed session.