Practical MCP with FastMCP & LangChain

Custom Storage Backends

The default state store keeps everything in memory on the same process that runs the server. That is perfectly fine for a single-server deployment where all requests go to the same process — which is the common case when using STDIO or a single HTTP server.

For deployments where multiple server instances share load — such as a containerised service behind a load balancer — in-memory state will not work, because two requests from the same client may land on different machines. FastMCP supports plugging in an external storage backend for these cases:

from key_value.aio.stores.redis import RedisStore

# All sessions share the same Redis store, so state survives
# across multiple server instances
mcp = FastMCP(