Model Context Protocol update makes MCP stateless for enterprise agents

The Model Context Protocol update released by the Agentic AI Foundation under the Linux Foundation is the largest revision to MCP since Anthropic introduced it 20 months ago. The release makes the protocol fully stateless, tightens authorization, sets a 12-month deprecation policy and promotes two agent capabilities to official extensions, changes its maintainers say are aimed at large enterprise deployments.

MCP is the open protocol used to connect AI agents with external software and services. The new release is not a financing event and carries no valuation, but the operational numbers are notable: David Soria Parra, an MCP co-creator and Anthropic maintainer, said SDK downloads have reached about 250 million a week, roughly doubling over six months.

What changed in the Model Context Protocol update?

• MCP now uses a fully stateless architecture, removing the need for sticky routing or shared session state between a client and one specific server instance.
• The authorization specification now requires validation of the issuer, or iss, parameter to guard against OAuth mix-up attacks.
• A formal policy gives developers at least 12 months between a feature being deprecated and the earliest date it can be removed.
• MCP Apps and MCP Tasks are now official extensions. Apps supports interactive server-rendered interfaces, while Tasks supports long-running asynchronous work.
• Multi-round-trip requests allow a client and server to exchange information during a single logical operation instead of treating every tool call as one pass.

Why does stateless MCP matter for enterprise agents?

The old MCP design required an AI application to keep a session tied to a particular server. In cloud environments where Kubernetes pods and other compute nodes appear and disappear behind load balancers, that made production systems fragile and harder to scale.

Den Delimarsky, a lead maintainer, said the previous setup forced teams to manage session stores and session IDs, and that requests could fail if the compute pod holding the session went down. Mazin Gilbert, executive director of the AAIF, said he has seen companies deploying tens of thousands of agents and argued that scale required this architectural shift.

The trade-off is that some state moves across the wire, making payloads larger. Soria Parra said those payloads are compressible and remain small compared with typical web HTTP traffic. Some lightly used capabilities are also being narrowed or removed, including out-of-band server logging, which Soria Parra said the team found had very little use after checking public GitHub activity.

How does the MCP update change security and governance?

The authorization changes align MCP more closely with OAuth 2.0 and OpenID Connect deployments. Delimarsky said the issuer-validation change was preventive and was not a response to active exploitation or a known breach.

The release also includes an Enterprise Managed Authorization extension developed with Okta. According to Delimarsky, the goal is to let companies make their corporate identity provider the authority for MCP server access, rather than allowing employees to connect with personal credentials.

The 12-month deprecation policy is aimed at enterprise buyers that need stable specifications before committing infrastructure. Delimarsky said maintainers consulted companies including Google, Microsoft and Amazon before settling on that window.

Governance remains a live issue because Anthropic created MCP and later donated it to the AAIF. Soria Parra said he technically has veto rights as lead maintainer and an Anthropic employee, while adding that he has not used them in protocol debates. The core maintainer group now includes people from Anthropic, Microsoft, OpenAI, Google and Amazon, and Gilbert said AAIF membership has grown from about 40 organizations at launch to 240.

What do Apps and Tasks add for AI agents?

MCP Apps lets servers deliver interactive interfaces such as forms, dashboards and visualizations inside AI clients. MCP Tasks gives servers a durable task handle for jobs that take time, so clients can disconnect or restart and later check status instead of keeping a long-lived connection open.

The next test is adoption on the new specification. The maintainers said success will be measured by server counts, working-group feedback and whether major contributors such as Microsoft and Google ship against the revised protocol.

This story draws on original reporting from VentureBeat.