Every multi-tenant MCP deployment starts with the same instinct: put a tenant_id in the JWT, scope every tool call to it, and call the isolation problem solved. It isn't. The gap between "scoped by claim" and "actually isolated" is exactly where these deployments fail in production β not in testing, where a single code path is exercised, but in production, where a new tool handler forgets the filter, a cache key drops the tenant dimension, or an agent legitimately holds a valid token but aggregates a result that happened to include data from a neighboring tenant.
This article is about what real isolation looks like across three architectural patterns, which layer each one enforces, and how to test the boundary rather than just build it.
Three patterns, not one
Multi-tenant MCP deployments settle into three recognizable shapes.β¦