MCP At Scale FAQ

Token bloat in MCP happens when too many tool definitions (schemas + descriptions) are loaded into the model context, inflating prompt size before reasoning starts.

Use search-based tool discovery, rank to a shortlist, enforce context budgets, filter by policy (role/tenant), and expand schemas on demand (summaries first, full schema later).

A context budget is a hard limit on tool/schema tokens per request. It keeps cost and latency predictable and prevents runaway prompts.

Lazy loading means loading tool schemas only after the agent identifies relevant tools (usually via search/ranking), rather than preloading the entire catalog into context.

Tool discovery is querying a registry (or authority layer) to retrieve a small, ranked set of candidate tools for a task, then loading only those tools into context.

An MCP registry is a catalog of tools/servers and their metadata used for discovery, versioning, ownership, and governance.

Shadow MCP is ungoverned MCP tooling that appears without centralized discovery, policy enforcement, or auditing. It often shows up first as tool sprawl and token bloat.