If your AI agents are useful, they are likely dangerous. As companies rush to connect agents to internal tools via the Model Context Protocol (MCP), we are seeing a classic infrastructure failure: adoption is sprinting ahead of governance. Nolan Sullivan, Founding Growth Engineer at Speakeasy, argues that MCP servers must now be treated as owned assets with a defined lifecycle, not just ad-hoc configurations buried in local JSON files. Without a structured approach to discovery, approval, and monitoring, organizations are accumulating 'shadow MCP'β€”unapproved servers that bypass security controls and create audit nightmares.

The Six-Stage Lifecycle for MCP Assets

Treating MCP servers like software products requires a closed-loop process. Sullivan outlines six critical stages: Discover, Review and Approve, Promote, Distribute, Monitor, and Update or Retire. Discovery is the hardest part because shadow servers, by definition, fall outside the approved path. Relying solely on network telemetry often misses local stdio servers or agent-side hooks. The solution is an immutable MCP inventory that records every server, sanctioned or not, including its URL, transport, credential type, and last-seen time. This inventory is the single source of truth that prevents the 'one-time scan' mistake, which quickly becomes stale as new agents spin up.

Why Gateways Often Fail at Discovery

A common misconception is that an MCP gateway solves governance. Most commercial and open-source gateways, such as Microsoft MCP Gateway or IBM ContextForge, start at the approved registry. They enforce rules on traffic they already know about, but they are blind to servers that bypass them entirely. Sullivan notes that a gateway without discovery only governs the servers that were already behaving well. True governance requires sensors where agents runβ€”such as device agents pushed through mobile device management or client configuration scansβ€”to catch the unapproved calls before they become security incidents. If your gateway doesn't see the traffic, it can't log it, and you can't audit it.

Closing the Loop: From Blocked Calls to Approved Access

Governance fails when it becomes a bottleneck. The goal is a closed loop where the output of one stage feeds the next. When an agent calls an unapproved server, the system should block the call but immediately offer a request path that lands on the server's inventory record for review. Once approved, the server should be automatically provisioned to the user's role via identity provider groups, eliminating manual ticket queues. This speed is critical; if the turnaround time for approving a new MCP server takes weeks, engineers will revert to wiring servers up themselves, perpetuating the cycle of shadow MCP. The metric that matters is the time between a developer wanting a tool and having it securely provisioned.

Key Takeaways

  • Inventory First: You cannot secure what you haven't found. Start with observe-only discovery to map all active MCP servers before enforcing blocks.
  • Role-Based Distribution: Avoid per-laptop config edits. Map MCP server access directly to identity provider groups to scale provisioning.
  • Audit Everything: Log every tool call with user, agent, server, and decision data. This log drives retirement decisions for unused servers.
  • Gateway Blind Spots: Standard MCP gateways do not automatically discover shadow servers; you need agent-side instrumentation or network sensors.
  • Retire Actively: Revoke credentials and remove servers from roles when they are unused or fail re-review, don't just delete the registry entry.

The Bottom Line

MCP governance isn't about stopping innovation; it's about making the safe path the easy path. If your platform team is drowning in manual server requests, you don't have a governance problem, you have a provisioning problem. Build the loop, or watch your shadow MCP footprint grow until it becomes impossible to audit.