MCP 2026: the protocol connecting AI agents to your business

On July 28, 2026, the team behind the Model Context Protocol (MCP) — the open standard that lets an AI agent read and act on a company’s tools: its CRM, its ERP, its invoicing, its spreadsheets — published the 2026-07-28 revision, described by several independent write-ups as the biggest update to the protocol since it launched, as reported by the official MCP blog and Google’s developer blog. It won’t make the mainstream press, but if your business uses or is evaluating AI agents wired into its systems — increasingly the norm in software built today — this update decides whether that connection survives production or stays a nice demo.

What MCP is, and why you’ve started running into it

Until recently, connecting an AI assistant to an internal system — to check stock, open a ticket, or summarize a folder of contracts — meant a custom integration for every model-and-tool combination. MCP was built as the “universal plug” to fix that: one standard that any model (Claude, GPT, Gemini) can speak to any system that implements it, without rewriting the integration each time. It’s the technical piece that lets a coding agent inspect your database, or a customer-support assistant look up a real order instead of making one up.

What changes with the July 28 revision

No more technical sessions: more stable, cheaper to run

The previous version of MCP kept an open session between the agent and each connected system, tracked with an identifier header (Mcp-Session-Id) that every server had to hold onto. The new revision removes that session entirely: every request now carries everything it needs, self-contained. In practice, a provider no longer needs special infrastructure to “remember” where a conversation was — requests can be routed to any available server, which cuts cost and adds stability to the part of the bill that pays for these integrations to stay reliable.

MCP Apps: dashboards and forms, not just text

The change with the most direct impact for a small or mid-sized business is the official MCP Apps extension: until now, an agent wired into your systems could only answer in text. With MCP Apps it can show dashboards, forms, and multi-step workflows right inside the conversation, rendered in a sandboxed environment for security. In practice: a finance team reviews a dashboard, not a block of JSON; an operations lead approves a change with a button, not a typed command. It’s the detail that decides whether an AI integration can be used by an entire team, or only by whoever knows how to code.

Every action, logged

One point worth flagging for anyone already dealing with AI Act transparency obligations: any action triggered from one of these panels goes through the same audit and consent path as a regular tool call. A click inside the interface isn’t a side channel that bypasses oversight — it’s logged exactly like any other action the agent takes.

Why this matters if your business uses (or is considering) AI agents

If your business already has an assistant wired into an internal system — or you’re weighing commissioning one — this update marks the difference between an integration that survives a two-user pilot and one that holds up when the whole staff uses it, without falling over or blowing up your server bill. And the jump from “text only” to “dashboards anyone can read” is probably the single change that most speeds up an AI agent going from a toy for the technical team to a tool the operations or support department actually uses without friction.

What to ask your software or AI vendor

1. Whether their AI integrations run on MCP, and if so, which version — the July revision is recent, and not every vendor has adopted it yet.
2. How every action an agent takes on your systems gets logged, especially when it touches customer data.
3. Whether your team sees plain text or dashboards and forms — this decides whether the whole team can adopt the tool, or only the people comfortable with commands.
4. What happens if they switch standards or model providers — an integration built properly on an open standard shouldn’t lock you into a single vendor.

How we handle this at Evicron

At Evicron, an AI and custom software studio based in Barcelona, we build our projects’ AI agent integrations on open standards like MCP specifically to avoid lock-in with a single model provider, and so that every action is logged from day one — not bolted on later when a client or an auditor asks for it. In our applied AI service, we start from which systems the agent needs to touch and under what level of supervision, before writing a single line of integration; in custom software, it’s the foundation we build any multi-system automation on.

The bottom line

On July 28, 2026, MCP — the standard connecting AI agents to a company’s systems — shipped its biggest revision to date: it drops the technical session to scale better, adds dashboards and forms so non-technical staff can use it, and audits every action with the same rigor. If your business has, or is evaluating, an AI agent wired into its systems, now’s the time to ask your vendor whether they’ve adopted it.

Want us to review how your business’s AI integration is built — or should be built? Get in touch: the first discovery session is free, and we reply within 24 hours.