While vendor agent products keep changing, one piece is moving the other way: the layer connecting AI models to enterprise systems is settling into a shared standard.
Out of one company's hands
In December 2025 Anthropic donated the Model Context Protocol to the Agentic AI Foundation, a fund under the Linux Foundation. OpenAI and Block joined as co-founders, alongside AWS, Google, Microsoft, Cloudflare, GitHub and Bloomberg.
That governance detail matters more than it looks. A protocol owned by one company can change direction to suit that company's commercial interests. Under community governance, decisions go through an open process, and organisations building on it have grounds to believe it will not disappear in the next product reshuffle.
On usage, WorkOS cites roughly 97 million monthly downloads of the Python and TypeScript SDKs, with support from Slack, GitHub, Google, Salesforce, Stripe, HubSpot, Shopify, Linear and Figma.
The July specification drops session state
On 28 July 2026 the new MCP specification moved the protocol to a stateless model. Previously each connection was tied to a session on a particular server instance; now each request stands alone.
It sounds technical and the consequences are very practical. No sessions means it runs behind an ordinary load balancer, scales by adding machines, and one server failing does not drop a user's session. This is precisely what something needs in order to move from prototype to production.
The release also adds multi-round-trip requests for steps that need confirmation, header-based routing, cacheable list results, and tighter authorisation aligned more closely with OAuth.
But it is not finished
The 2026 roadmap summarised by WorkOS lists four remaining enterprise gaps: no standardised audit trail, no multi-tenancy model, no answer for rate limiting and cost attribution for autonomous agents, and no portable configuration across clients.
All four are questions an IT department raises the moment a project leaves the lab. So MCP is mature enough to build on, while the governance layer remains yours to provide.
For enterprises in Vietnam
The practical significance is architectural. If you intend to let AI read data from an ERP, an MES or a document management system, the question of how to connect had no common answer — every vendor did it differently, and changing supplier meant rewriting.
Now there is one way to connect that many parties support. A company writes an MCP server for its internal system once, and it works with different models. Change model vendor and the connection layer stays. That is a concrete way to reduce lock-in rather than a slogan.
But do not read it as "open MCP for every system". An MCP server is a door into internal data. Because the standard still lacks an audit trail and a multi-tenant permission model, that part is yours to build: who may call what, where it is recorded, what the limits are.
A sensible start is one system, read-only, for a small group, with full logging from day one. Once you can see what the agents call and how often, expanding has a basis.
Sources
Model Context Protocol Blog — The 2026-07-28 Specification: https://blog.modelcontextprotocol.io/posts/2026-07-28/
WorkOS — Everything your team needs to know about MCP in 2026: https://workos.com/blog/everything-your-team-needs-to-know-about-mcp-in-2026