MCP goes stateless: the spec finally clears the last enterprise blocker

In plain terms. The Model Context Protocol - the spec that lets AI assistants call external tools - just shipped a revision that removes the requirement for a stateful connection between client and server. That is the change enterprise infra teams have been waiting for.

Why stateful was a problem

The original MCP transport assumed a persistent client-server session. If your assistant needed a tool call, the assistant and the tool shared a live channel for the duration of the interaction. Fine on a laptop; awkward at scale.

The awkwardness surfaces as soon as you put a load balancer in front of it. A stateful session means every subsequent call has to land on the same backend instance. That kills the cheapest deployment pattern in enterprise infra - spray requests across a stateless pool, autoscale on load, tolerate node loss. You end up building session-affinity kludges, sticky routing, or worse, connection managers that become their own SPOFs.

What the new spec actually changes

Per the reporting, the revised transport lets servers advertise a stateless mode. In practice:

# before (paraphrased)
POST /mcp/session → session_id
POST /mcp/call { session_id, tool, args } # must hit same node
POST /mcp/close { session_id }

# after - stateless mode
POST /mcp/call { tool, args, ctx } # any node, no session

The ctx is the piece that carries whatever the server needs about prior turns, pushed by the client. It's the same pattern REST APIs settled on twenty years ago - statelessness bought with slightly larger requests.

Under the hood: the trade-off you now own

Stateless is not free. Two things move onto the client:

• Context management. The client is now the source of truth for any per-conversation state a tool needs. Bigger payloads, and a new class of bug: forgetting to propagate.
• Idempotency. Once any node can serve any call, retries become plausible in more places. Tool authors need to make their calls idempotent or deal with double execution.

For most enterprise deployments this trade is worth it. For low-latency assistants where round-trip bytes matter, it's a real cost.

What this signals

The MCP ecosystem thread has been about the plumbing maturing from "prototype-quality" to "we can actually put this behind an SRE team." Stateless transport is the concrete step in that direction. It also tells you where the spec authors think the growth is: inside enterprises, behind proxies, not on individual developer machines.

So what

For a builder: if you shipped an MCP integration and hit session-affinity pain, the fix is now spec-blessed rather than bespoke. For a decider: MCP just became procurement-friendly. Expect vendors to advertise "MCP-native" the way they used to advertise REST-native.