When MCP first launched, I was excited enough to build an early one and test it out. For anyone outside the AI bubble, MCP is basically a standard way for AI systems like Claude to connect to the tools, data and systems inside a business. In theory, that is a very big deal.
Instead of AI just answering questions from memory, it can look things up, pull information from real systems, take actions, follow business rules and work with live context.
That was the promise.
The original promise, and the messy reality
When I first built with it, the reality was messier. The spec was ahead of the implementation. Not every client had implemented all parts of the protocol. Some things were brittle. Some things worked in one place but not another. The benefits were obvious, but they could not be fully realised yet.
Then Claude Code happened
For me, and apparently half the internet, Claude Code was the moment agentic AI became real. Not "agents" as a slideware concept. Actual agents.
Inspect the context. Make a plan. Use tools. Run commands. Check the result. Recover from mistakes. Keep going.
And ironically, that made me more sceptical of MCP for a while. Claude Code made the command line feel like the native home for agents. Developers already had CLIs. Scripts. Local tools. Git. Tests. Logs. Databases. Build systems. The shell was already full of useful tools. Claude Code was smart enough to use them.
So a lot of people, me included, started thinking: do we really need MCP for this? For developer workflows, maybe not always. There was a lot of anti-MCP energy for a while. I understood it. I probably agreed with a lot of it.
But I've changed my mind.
Then Claude Cowork happened
Claude Code was the developer moment. Claude Cowork feels like the knowledge work moment.
The agent is moving from software development into general business processes: finance, compliance, operations, renewals, reporting, customer work, research, internal admin and document-heavy workflows.
And that world is different. Most businesses do not run on a clean command line. They run on email, spreadsheets, PDFs, CRMs, finance systems, document stores, approval processes, messy databases, legacy apps and a lot of tribal knowledge.
In that world, the model needs more than a prompt. It needs permissioned access to live systems. Governed access to private data. Domain-specific actions. Approvals, constraints, memory, evidence and audit trails. It needs to know not just what it can do, but what it is allowed to do.
From "can it call a tool?" to "can it run a business?"
That is where remote MCP starts to matter.
The first wave of MCP was about: can the model call a tool? The next wave is about: can the model understand the business well enough to use the right tools, in the right sequence, with the right guardrails?
Why thin AI wrappers are getting hard to defend
That also changes how I think about AI wrappers. A thin AI wrapper around Claude is getting harder and harder to defend.
For most businesses that look beyond Copilot, Claude is going to own the chat UI. (Copilot is the Windows of AI - the default many organisations will land on simply because it is already there.) For everyone else, that battle is basically over. There is not much point building another standalone app that is mostly a chat window in front of the same frontier model.
But an MCP server that gives Claude deep access to a regulated workflow, high-quality domain data, permissions, evidence, auditability and useful business actions? That feels like a real place to build.
Thomson Reuters announced exactly this move in May 2026. The next generation of CoCounsel Legal is being rebuilt on Claude's Agent SDK, and Claude connects to CoCounsel through MCP. A lawyer can sit inside Claude, pass a task to CoCounsel's workflows, and get back work grounded in 1.9 billion Westlaw and Practical Law documents with citations validated by KeyCite. Thomson Reuters calls it "fiduciary-grade" AI. The chat UI is Claude. The depth, the authoritative content and the professional accountability are Thomson Reuters. That is what the layer underneath looks like in a regulated profession.
The opportunity is not to build another chat UI on top of Claude. The opportunity is to build the layer underneath Claude that makes it useful inside a real business.
Long live MCP
I was sceptical for a while. I think I was too quick to dismiss it.
MCP was dead. Long live MCP.