The Protocol Powering AI Tools Is Burning Through Your Budget
MCP — the protocol that connects AI tools to external data — crossed 97 million monthly downloads. But it has a cost problem that's eating through AI budgets, and the industry is splitting over what to do about it.
There's a good chance you've never heard of MCP. But if your business uses AI tools that connect to databases, file systems, calendars, or any external service, MCP is probably running behind the scenes.
Model Context Protocol is a standard that lets AI systems talk to external tools. We covered the ecosystem in The MCP Ecosystem Just Got Real — it's growing fast. Think of it as the plumbing that connects your AI assistant to your CRM, your calendar, your documents. Every major AI provider supports it — Anthropic, OpenAI, Google, Microsoft, Amazon. It crossed 97 million monthly SDK downloads in February 2026.
It's also quietly burning through AI budgets in ways most businesses don't realize.
The Cost Problem Nobody Warned You About
AI tools charge by usage — specifically by "tokens," which are chunks of text the AI processes. Every word you send and every word it sends back costs tokens. When you're paying per token, efficiency matters.
Here's where MCP creates a problem. Every time your AI tool starts working, MCP loads the descriptions of every connected tool into the AI's context. Not just the tools it needs for the current task — all of them. Every tool, every parameter, every description.
At the Ask 2026 conference in March, Perplexity's CTO shared a real deployment where three MCP servers consumed 143,000 tokens out of a 200,000 token context window — before the AI even started working on the actual task. That's 72% of the available capacity used up by plumbing.
For businesses, this translates directly to money. More tokens consumed means higher API bills. It also means the AI has less room to work with your actual data and instructions, which can reduce the quality of results.
The Industry Is Splitting
The MCP situation has created an unusual divide in the AI industry.
On one side, adoption is surging. Enterprises are building integrations. Google released a Colab MCP server that lets AI agents access GPU computing remotely. New MCP servers appear daily for everything from fraud prevention to database management.
On the other side, prominent voices are walking away. Perplexity publicly announced they're moving away from MCP internally. Y Combinator's CEO chose custom CLIs instead. A tool called mcp2cli hit the front page of Hacker News by demonstrating that it could reduce tool description costs from 121 tokens per tool to 16 — a 96-99% savings.
What Anthropic Is Doing About It
In March, the team behind MCP published a 2026 roadmap with four priorities:
Streamable HTTP transport — making MCP work better at scale for enterprise deployments.
Tasks primitive — letting AI agents manage long-running operations instead of single requests.
Enterprise readiness — audit trails, SSO, and the kind of security features large companies require.
Standard metadata format — this is the interesting one. A way for tools to describe their capabilities without loading full descriptions into the AI's context. If it works, it could solve the token waste problem at the protocol level.
That last point matters. If MCP can describe tools in a compact format that AI systems understand without consuming thousands of tokens, the cost problem shrinks significantly.
What This Means for Your Business
If you're using AI tools through a vendor, you probably don't interact with MCP directly. But you're paying for it — through higher API costs, slower responses, or reduced quality when the AI's context window gets crowded.
A few practical things:
Ask your vendor how many MCP servers they run. As we detailed in Why Your AI Agent Costs 10x What It Should, more isn't always better. Each additional connection adds token overhead.
Watch for "context window" issues. If your AI tool gives shorter or less detailed responses than it used to, context crowding from MCP tools might be the cause.
Monitor your AI costs over time. If usage charges are creeping up without a clear reason, bloated tool descriptions could be a factor.
The protocol itself isn't bad — it solved a real problem by standardizing how AI tools connect to external data. But like a lot of infrastructure decisions, the costs show up later and in places you don't expect.
The 2026 roadmap suggests fixes are coming. Until then, the plumbing is expensive.
Related: The MCP Ecosystem Just Got Real and Why Your AI Agent Costs 10x What It Should.
Stay Connected
Get practical insights on using AI and automation to grow your business. No fluff.