Speakeasy Founding Growth Engineer Nolan Sullivan has published a guide defining the "MCP inventory" as a continuously updated record of every Model Context Protocol server within an organization. The article clarifies that this inventory tracks server identity, exposed tools, authentication credentials, ownership, and sanctioned status. Sullivan distinguishes this from an "agent inventory," which records the AI agents calling those servers, noting that both records are necessary to enforce security policies effectively.

The Blind Spot in Traditional Asset Management

Sullivan points out that standard endpoint scans and configuration management databases (CMDBs) often fail to capture MCP servers because they are frequently configured as simple JSON entries in dotfiles, such as ~/.cursor/mcp.json. For example, an engineer’s laptop might list Cursor in the software register, but miss a GitHub MCP server configured locally that uses a personal access token with broad read/write repository permissions. This gap means security teams lack visibility into the actual tools and credentials their AI agents are utilizing.

Distinguishing the Public Registry from Internal Governance

The guide emphasizes that the public MCP Registry at registry.modelcontextprotocol.io serves as a catalog for discovery, not a governance tool for individual enterprises. An enterprise MCP inventory must track servers that are actually running within the organization, including local stdio servers, remote SaaS URLs, and internal generated servers. Sullivan argues that while the public registry supplies candidates for approval, the internal inventory records which of those candidates are actively being called by agents and under what authentication constraints.

Essential Data Fields for a Live Inventory

To be effective, an MCP inventory record must include specific fields: server identity (URL or package specifier), transport type (stdio or remote), tool disposition (read-only vs. destructive), credential method, human owner, connected agents, sanctioned status, and last seen timestamp. Sullivan notes that tool lists are dynamic, requiring a live feed rather than a static snapshot. A dormant server without recent activity is flagged as a revocation candidate, while an active, unsanctioned server triggers immediate investigation.

Enforcing Least Privilege Through Gateway Integration

The article details how Speakeasy’s AI Control Plane maintains this inventory using two live feeds: endpoint discovery hooks that report servers connected during agent sessions, and a gateway that records traffic from sanctioned servers. This architecture allows for granular policy enforcement, such as granting an agent access to get_invoice while blocking create_refund. Sullivan explains that when a server is compromised or "rug-pulled," the inventory serves as the central point for revoking access and providing audit evidence.

Key Takeaways

  • MCP inventories track servers and tools, while agent inventories track the callers; both must link for effective security.
  • Traditional asset scans miss MCP servers configured in dotfiles or remote URLs, creating significant security blind spots.
  • The public MCP Registry is a discovery catalog, not a substitute for an internal, continuously updated inventory.
  • Live inventories require endpoint hooks and gateway traffic monitoring to remain accurate, as static spreadsheets decay quickly.
  • Security policies rely on the inventory to enforce least privilege at the tool level and manage credential revocation.

The Bottom Line

Asset scans are obsolete for the AI agent era because they miss the server-side connections that actually execute code and access data. If you aren't tracking MCP servers in a live inventory, you aren't securing your AI agents; you're just guessing at their reach.