Anthropic, Cloudflare, AWS, Microsoft — each has shipped a cheaper, faster way to connect AI agents to your systems. The MCP standard itself survives. Its position in the architecture does not. What was sold as the universal interface for AI agents is being demoted to plumbing — still useful, but no longer the place where decisions get made. The shift happened in under twelve months, this week we explore why.
Focus On: The Token Bill That Forced the Industry’s Hand
The arithmetic is brutal. A single GitHub connection through MCP loads 93 separate command definitions into an AI agent’s working memory — 55,000 tokens of overhead before any work gets done. (Tokens are the units AI vendors meter and bill against — every cost projection in your AI budget eventually traces back to them.) Connect three common business tools — GitHub, Slack, Sentry — and 143,000 of an agent’s 200,000-token working memory is gone before it reads its first instruction. Cloudflare’s published benchmark shows the ceiling: their full developer platform exposed through MCP consumes more than 1.17 million tokens per request. The replacement pattern reduces that to roughly 1,000. A 99.9% reduction.
Anthropic’s engineering team published the canonical case study in November. A workflow that consumed 150,000 tokens under MCP dropped to 2,000 under the new approach. Accuracy improved at the same time — agent success rates climbed from 49% to 74% on identical evaluations. Independent benchmarks from Speakeasy and Scalekit confirm the order of magnitude. Thirty-two times cheaper. Sometimes a hundred.
The procurement question this raises is the one I rarely hear in vendor reviews. What does the connection layer cost in tokens, at our scale, on our actual workflows? Most enterprise AI budgets price the model and ignore the connection overhead entirely. That overhead is now the line item driving the bills nobody warned the CFO about.
The Category Error CIOs Are Still Making
Forrester’s framing is the clean one. MCP is wiring. It moves data between AI agents and the systems your business already runs. Wiring is not governance — the rules about who can do what, with which data, under whose authority. CIOs who treated MCP as a control layer are discovering the control layer has to be procured separately, urgently, and from a different vendor. Cloudflare, Strata Identity, AWS Bedrock AgentCore and a half-dozen specialised gateways are positioning into that gap.
Clutch Security research, surfaced by Strata Identity in January, found the average 10,000-person organisation already runs 3,056 unsanctioned MCP connections — 15% of employees with two each, on average. None governed by enterprise identity systems. Most using personal credentials granting far more permission than the user actually needs. Your CISO doesn’t know they exist. Your finance team is paying for the tokens. That’s not a future risk. That’s last quarter’s invoice.
What This Means at Your Next Vendor Review
The move is straightforward, even if most enterprise teams haven’t made it yet. Treat MCP as backend wiring. Buy the cheaper, sandboxed alternatives every major vendor now ships. Procure governance separately, from a vendor whose actual business is governance — not from the AI platform that sold you the wiring in the first place. Stop signing agent contracts whose pricing model conveniently leaves out the connection overhead.
The vendor signal to watch is who already ships the cheaper pattern by default. Cloudflare’s own MCP server explicitly warns you the old approach costs 244 times more. Anthropic’s Tool Search Tool defers loading commands until the agent actually needs them — cutting cost and improving accuracy in the same move. The vendors who haven’t shipped equivalents are the ones who’ll need to explain why on their next earnings call.
There is a single question worth taking into your next AI vendor evaluation. Did anyone in the room ask what the connection layer would cost, at our scale? If the answer is no, what you signed was a budget approval, not an architecture choice. The two are not the same thing.
Follow me
That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.
Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.