Anthropic updates MCP with stateless core, stronger security and task support
New Delhi: Anthropic has announced the fifth specification release of the , introducing changes aimed at making AI tool connections easier to deploy and scale. The MCP 2026-07-28 specification shifts the protocol to a stateless request and response model, strengthens authorisation and moves key features into an official extensions system.

MCP is an open standard that lets AI applications connect with external tools, business software and data sources. Anthropic said the protocol has crossed 400 million monthly software development kit downloads. Claude’s connectors directory now lists more than 950 MCP servers, according to the company.
MCP moves to a stateless model
The biggest change is the removal of . Earlier versions relied on an initial connection and session information that servers had to maintain. Under the new model, every request carries the details needed to process it.
This means a request can reach any available server behind a normal load balancer. Developers will no longer need shared session storage simply to keep an MCP connection working. The change could make MCP servers easier to run on cloud, edge and serverless infrastructure.
The old initialize and initialized exchange has been retired, along with the Mcp-Session-Id header. Clients that need to check a server’s features can use an optional server/discover request.
Applications can still remember information between calls. A server can create an explicit handle and ask the AI model to send it back during the next tool call.
New way to handle questions during a task
MCP has introduced Multi Round-Trip Requests, called MRTR. It replaces several server-initiated requests that previously needed a connection to remain open.
For example, an AI agent booking a meeting may need the user to confirm a time. The server can return an input_required result containing the question. The client then sends the original request again with the user’s answer attached.
Better routing, caching and authorisation
MCP requests now include method and tool names in the Mcp-Method and Mcp-Name HTTP headers. Gateways and security systems can use these headers to route requests, apply access rules or measure usage without reading the full request body.
Tool, prompt and resource lists can carry cache details such as how long a result remains valid. This can reduce repeated requests and keep tool lists in a stable order.
The specification adds issuer checks based on RFC 9207 and binds client credentials to the authorisation server that issued them. Dynamic Client Registration is being deprecated in favour of Client ID Metadata Documents, though older implementations will continue working for now.
Tasks and MCP Apps
Tasks have moved out of the experimental core into an official extension. They support longer operations that may continue after the first request, with options to check progress or receive updates.
MCP Apps, enterprise-managed authorisation and Tasks now sit inside a versioned extensions framework. This gives developers a defined route to add interactive screens, identity controls and longer workflows without changing the base protocol.
Roots, Sampling, Logging and the older HTTP plus Server-Sent Events transport have entered deprecation. The MCP team said affected features will continue working for at least 12 months, giving developers time to migrate. TypeScript, Python, Go and C# software development kits are being updated for the release.
