Scale an MCP Server

With MCP 2026-07-28, protocol requests are self-describing and no longer require sticky protocol sessions. That allows ordinary horizontal scaling.

Keep instances disposable

Application jobs, user records, approvals, and durable task state belong in shared stores rather than in one process's RAM.

Scale on useful metrics

• Request concurrency.
• Tool latency.
• CPU/memory.
• Downstream pool saturation.
• Queue depth for async work.

Protect downstream systems

Adding ten MCP replicas does not create ten times the database capacity. Use pool limits, concurrency controls, and backpressure.