All articles

AI & Product Engineering

MCP in 2026: What Product Teams Should Know Before Connecting AI Agents to Their Tools

The latest Model Context Protocol release adds a stateless core, clearer routing, cache hints, and authorization changes. Here’s how to think about it in a real product.

A central AI tool connected to application, database, and service icons

AI agents become more useful when they can access the tools and data a person already uses. The Model Context Protocol (MCP) is an open way for AI applications to discover and call those tools. Its July 28, 2026 specification update adds production focused changes, including a stateless protocol core, header based routing, cache hints, authorization hardening, and a formal extensions framework.

MCP is an integration layer, not a product requirement

The 2026-07-28 release removes protocol level sessions for HTTP requests. Each request carries its own context, which lets ordinary infrastructure route requests without pinning a client to one server instance. The update also adds method and tool names to headers so gateways can route and meter calls, plus cache hints for tool and resource lists. The official MCP 2026-07-28 specification announcement

This can make integrations easier to operate, but MCP does not replace your application’s business logic, authentication model, or API design. If your product already has reliable APIs, an MCP server can act as a carefully scoped adapter for AI clients that need to discover and use selected capabilities.

Design tools around safe, understandable actions

Expose focused operations with clear input schemas. A tool such as “find available appointments” is easier to reason about than a general purpose database query. Keep read and write actions distinct, validate every argument on the server, and return structured results that tell the model what happened and what it can do next.

The new specification’s stateless transport does not mean every application must be stateless. If a task needs to carry state between calls, the MCP announcement recommends making that state explicit through an application handle passed as a tool argument. This keeps state visible and lets the application decide how it is stored and authorized. MCP’s explanation of explicit state handles

Make identity and permissions part of the integration

An agent should not receive a broader set of permissions than the user or task needs. Bind access to the right account and resources, validate authorization on every operation, and log which actor invoked each tool. Ask for human confirmation before actions that create significant consequences, such as sending money, changing access, or submitting a final transaction.

MCP’s July release includes authorization hardening and updates to how clients validate authorization servers. Teams implementing the new version should read the current specification and the migration notes for their chosen SDK rather than copying an older example unchanged. MCP TypeScript SDK v2 documentation

Build observability and recovery before adding more tools

Plan for timeouts, retries, duplicate calls, rate limits, and partial failures. Make write operations idempotent where possible, report errors in a way the client can act on, and keep an audit trail for both successful and rejected requests. Monitor which tools are called and where users abandon a workflow, while protecting private data in logs.

Start with one narrow workflow and a small set of tools. Test it with realistic user requests, permission boundaries, and failure cases. Expand only after the integration behaves predictably and the product team can explain how a call is authorized, monitored, and reversed when appropriate.

Apply this to your product

Start with one user journey or integration in your current product. Document the expected behavior, review the linked source material, and agree on the checks needed before release. The right implementation depends on your existing architecture and operational requirements.

Further reading

MORE FROM OUR TEAM

Related Engineering Insights

All articles

READY WHEN YOU ARE

Have a question about your next build?

We’d be glad to talk through your goals and possible next steps.