Claude daily update: MCP 2026-07-28 makes integrations easier to operate

MCP 2026-07-28 is now stable, and Anthropic is rolling out support across Claude products. The big change is easy to miss behind the protocol language: every request should be able to stand on its own. At the same time, the expanded Anthropic and Cognizant partnership shows how Claude is moving into real operating workflows where specifications and evaluation stay attached through production.

This is not a new Claude Code version. It is a broader Claude update about integrations that should be easier to operate and review.

Sources: Bringing MCP 2026-07-28 to Claude. Cognizant and Anthropic expand their partnership.

Claude daily update: MCP leaves session handling to the application

MCP, the Model Context Protocol, is an open standard that connects AI applications to external tools and data sources. In version 2026-07-28, protocol-level sessions and the old client-server initialization exchange are gone. Each request carries the protocol version and client capabilities instead. A server can describe its support through the new server/discover call.

That makes MCP servers easier to run behind ordinary load balancers, on serverless platforms, and at the edge. A dropped connection no longer has to leave a hidden protocol session behind. If a workflow needs memory across calls, the server should create an explicit handle and return it as an ordinary tool argument.

This is a breaking specification change, not a switch every Claude user has already received. Anthropic says support is rolling out across Claude products, while MCP clients and servers adopt the version at their own pace.

Sources: MCP 2026-07-28 specification. Full MCP key changes.

Long-running Claude workflows get explicit tasks

MCP Tasks is an official extension for work that outlasts a normal tool call. A server can return a durable task ID, the client can recover status after a restart, and the workflow can pause in input_required when a person needs to answer or approve something.

That fits a Claude workflow which starts a data run, waits for document review, or follows an external job. Progress stays visible without keeping a network connection open. Support is optional and must exist on both the client and server.

Authorization has also been tightened for common OAuth 2.0 and OIDC environments. This fits Anthropic's work on centrally managed authorization for Claude connectors: an identity provider can scope access by group, while short-lived tokens and bounded roles follow the organization's normal access model.

Sources: MCP Tasks. Centrally manage authorization for MCP connectors.

Cognizant shows what Claude needs around it

The second signal in this coverage window is less technical but just as practical. Anthropic says more than 30,000 Cognizant associates have completed Claude training. In Cognizant's Flowsource platform, project specifications, coding standards, and architecture blueprints direct Claude Code, then its output is evaluated before production.

That is a better starting point than a prompt on its own. Connect Claude to real systems, but attach sources, access, acceptance criteria, and a clear review point to the assignment. Anthropic's examples include a contract workflow where Cognizant reports shorter review time and higher extraction accuracy in that deployment. Those results are not universal promises, but they show the kind of measurement that belongs beside the integration.

Source: Cognizant and Anthropic expand their partnership to bring Claude to enterprise clients.

Try this prompt this week

Human step: Choose one existing or planned Claude/MCP workflow. Put its architecture, authentication, and operating documents in the same workspace. First verify which MCP features your actual Claude client and server support.

Read the documents for the selected Claude/MCP integration.
Map session dependencies, identity, tools, data, and long-running tasks.
Flag what MCP 2026-07-28 changes and what still depends on client support.
Compare the map with the business specification and acceptance criteria.
Propose one reversible integration test without changing anything.
End with owner, access, review point, and success measure.

Good output should:

• Separate verified configuration from assumptions and pending product support.
• Show where state lives and how interrupted work resumes.
• Tie every tool and data source to a scoped permission.
• Give the test a measure a person can review after the run.

This is a natural Tool Forge job: build the connection around a real workflow, with identity, task status, and evaluation treated as parts of the integration itself.