Frequently Asked Questions
Common Questions About Deploying MCP in Production
Everything you need to know before registering.
What infrastructure is needed to run MCP in production? +
Production MCP infrastructure requires: containerised MCP servers (typically Docker) with health check endpoints that the orchestrator polls to verify availability, a service registry where agent servers register their addresses and capabilities, a load balancer for MCP servers that need to handle high request volume, persistent storage for the Glass-Box logs and episodic memory, and monitoring infrastructure that tracks tool invocation latency and error rates per agent server.
How do I containerise MCP agent servers for production deployment? +
Containerising MCP agent servers uses Docker with a Python base image, the MCP server dependencies installed from a pinned requirements file, health check configuration that verifies the server can accept tool invocations, and resource limit configuration that prevents any single server from consuming excessive compute. The workshop covers a Docker configuration pattern for MCP servers that works across development, staging, and production environments.
How do I implement health checks for production MCP servers? +
Production MCP server health checks verify that the server can accept and respond to tool invocations within an acceptable latency threshold. A simple health check tool that echoes a test payload is registered on each server and polled by the orchestrator's availability monitor. Servers that fail health checks are temporarily removed from the orchestrator's active server list until they recover, preventing the orchestrator from routing tasks to unavailable agents.
How do I perform zero-downtime updates to production MCP agent servers? +
Zero-downtime MCP server updates use a blue-green deployment pattern: the new server version is deployed and health-checked alongside the running version, the orchestrator's service registry is updated to route new requests to the updated server, running requests on the old server are allowed to complete, and the old server is shut down once its request queue is empty. This pattern requires the orchestrator to track which server version each active request was dispatched to.
What monitoring should I set up for production MCP deployments? +
Production MCP monitoring covers: tool invocation latency histograms per server and tool (to detect performance regressions), error rates by error type (to distinguish transient from persistent failures), connection pool utilization (to detect capacity issues), Glass-Box trace completeness (to verify no requests are dropping steps), and semantic drift metrics (to detect when agent behavior changes in ways not reflected in error rates).
How do I scale MCP agent servers horizontally in production? +
Horizontal scaling of MCP agent servers uses multiple instances behind a load balancer for stateless agents, consistent hashing for agents that maintain per-request state, and auto-scaling policies based on request queue depth and tool invocation latency. The workshop covers the stateless agent design pattern that makes horizontal scaling straightforward, and the session affinity patterns needed for the few agent types that require state continuity across tool invocations.
