Stateless MCP seen as step forward for enterprise AI
Experts say the project's update this week will make it more suitable for enterprise production use, but it comes with a breaking change for existing systems.
An open source protocol that facilitates AI agents' use of tools and data sets has already been widely embraced in the industry, but a change just this week signals a key step towards maturity for enterprise AI, experts said.
Model Context Protocol (MCP) originated at Anthropic and was donated last year to the Linux Foundation subsidiary Agentic AI Foundation. It defines a standard format used by MCP servers to expose tools, data and prompts to MCP hosts, or clients, which send requests on behalf of AI agents.
Originally, MCP servers and clients used "sticky sessions," which had to be explicitly enabled in code to communicate between clients and specific MCP server instances. As part of a major release this week, the project shifted to a stateless transport architecture, decoupling MCP clients from specific back-end MCP servers and enabling communication through load balancers with pools of servers, increasing scalability.
This architecture aligns more closely with how most enterprise applications operate, said Nate Barbettini, founder and a member of the technical staff at MCP gateway vendor Arcade.dev and an upstream MCP contributor, in an interview with TechTarget.
"The early design of MCP was heavily inspired by the Language Server Protocol, which is really only relevant when you're talking about something on your local machine," Barbettini said. "As MCP has grown up and [been used] not just on your local machine, but also remotely, and now over HTTP … that has forced some evolution in how we think about the protocol. I see it as a maturity point, a growing-up moment for MCP."
Removing the requirement to include specific server information in communications aligns with best practices for both enterprise AI performance and security, according to one IT pro who has experimented with creating his own MCP servers internally.
Suresh Gangula
"With state, there's a two-way authentication or two-way verification and every time, the request has to carry a particular session ID, and it has to do a handshake between the server and the client," said Suresh Gangula, DevOps engineer at a Fortune 100 company that he requested not be named because of policies prohibiting him from representing it in the press. "It's a burden for service to send it every time. So with the new stateless approach, it just carries an underscore meta field, which is more secure and faster."
An underscore meta field obscures specific server and session information from a potential attacker, who could otherwise use that data to gain access to the back-end system. Instead of having such data coded into requests, it could be supplied by a separate, more secure system, which is a better practice for enterprises, Gangula said.
MCP brushes up auth specs, adds core stability
Industry analysts hailed the release as a milestone for a standard that has come to comprise a large share of AI agent infrastructure, but that has required rapid development to reach full maturity.
"MCP is being redesigned for how modern web systems actually operate," said Rob Strechay, founder and principal at Smuget Consulting. "Removing protocol-level sessions means requests can be handled by any server instance behind an ordinary load balancer, eliminating the requirement for sticky routing and shared session stores and enabling horizontal scale-out instead of ever-larger scale-up servers."
The July 28 MCP release also updates the specification to be more explicitly in line with OAuth 2.0 and OpenID Connect authorization and authentication practices, which had been less explicitly worded in the past, Barbettini said.
"It's an important clarification," he said. "We kind of found it the hard way -- we operate a pretty big MCP deployment, and we saw a lot of bug reports from our customers early on. … We traced it back to the fact that some MCP clients didn't actually follow the exact strict guidance that's in, for example, the OpenID Connect spec or the OAuth spec around how you're supposed to handle some edge cases."
Growth, and growing pains
We have definitely seen some MCP sprawl, so this could create issues for organizations once the grace period ends.
Jim Mercer, Analyst, IDC
While industry watchers generally welcomed the move to a stateless communication approach between MCP clients and servers, the update represents a breaking change. This means that after a 12-month conversion period, communication could fail among clients and servers if some haven't been updated to the new version.
"Adopting the new version requires refactoring to take advantage of its benefits, and MCP providers would need to do so in the near term -- 12 months -- to implement the changes," said Jim Mercer, an analyst at IDC. "We have definitely seen some MCP sprawl, so this could create issues for organizations once the grace period ends."
There is also the potential for disruption and fragmentation across different versions of MCP clients and servers as developers make updates, Barbettini said.
"If companies have custom clients or custom agents, they need to be updated to speak the new protocol, and that's not trivial," he said. "It's a pretty significant change."
However, this week's release also introduces mechanisms intended to make future updates less disruptive: a formal lifecycle with at least 12 months between feature deprecation and removal, and optional extensions that contributors can develop independently of the core spec.
"Yes, it’s a hard break, but as both versions can coexist during the transition, and as making the adjustments to the server mainly requires a rebuild based on the new version of the SDK, the impact seems on the low side," said Torsten Volk, an analyst at Omdia, a division of Informa TechTarget. "Ok, you need to still fix potential budget or rate limits or some other small stuff that’s based on the server-side session, but that doesn’t sound daunting to me.
"Twelve-month transition windows are what one would expect from enterprise-grade software. That's a part of growing up for MCP."
That doesn't mean MCP doesn't have further growing to do, Volk said.
"To fully grow up, MCP also needs to support agent-native identity instead of the human user's delegated identity, as the agent should not necessarily be able to have all the same privileges of the human identity," Volk said. "Agents basically need to be able to show up at the MCP with a slice of their user's permissions and then receive their responses accordingly."
Beth Pariseau, senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism. Have a tip? Email her or connect on LinkedIn.
Suresh Gangula
