The Model Context Protocol ecosystem crossed 13,000 servers in May 2026. Every week brings new GitHub repos, new announcements, and breathless benchmarks comparing which AI coding agent installs the most servers. The narrative is growth, growth, growth—and nobody's putting the real numbers on the marketing slides.
The Context Tax Nobody Calculated
Here's what's actually happening when you connect an LLM to an MCP server: every tool call becomes a round trip through the protocol layer. The model gets a structured description of the tool. The tool runs. The result gets stuffed back into context. For something as simple as an ls command, you've added 500 to 2,000 tokens to your prompt window that weren't there before. That's not a bug—it's math. The author ran experiments across three production projects—a code review pipeline, an automated PR triage bot, and a documentation updater—connecting every "recommended" MCP server available: GitHub, Filesystem, Playwright, Slack, Linear, the whole stack. The results were consistent and uncomfortable. Code review costs jumped from $0.003 per PR to $0.11—a 37x increase. PR triage went from $0.02 daily to $0.38—19x more expensive. Documentation updates hit a 40x multiplier. MCP costs 10 to 32 times more tokens than a direct API call to the same tool.
Pick Three, Not Twelve
The practical advice after six months in production: choose three servers maximum per agent, and pick them based on task frequency, not capability breadth. The framework that works is one primary action server (GitHub for code review flows, a database MCP for data agents), one context server to prevent hallucinations (code search, knowledge base lookup), and one utility server for the boring stuff that makes the agent look competent (Filesystem, Slack). The temptation to add more is real—MCP servers are fun to configure, the ecosystem is impressive, and saying "I run 11 MCP servers" sounds more serious than "I run three." Resist it. Every server you add is a context tax paid on every single prompt that follows.
Security Surface Nobody's Auditing
Here's the second problem nobody's discussing in the growth metrics: MCP servers run with the permissions of the agent's environment. When you connect a filesystem server, you're not just giving Claude the ability to read files—you're giving whatever that server's runtime is permission to execute in your environment. This matters more as the ecosystem fragments. Of the 13,000+ MCP servers cataloged in mid-2026, governance transfer to the Linux Foundation's AAIF is recent. The security review process for community-maintained servers is still maturing. Some servers are single-developer projects with no audit history. Audit them the way you'd vet a dependency from an unknown maintainer: check permissions requested, check what the server actually does with them, then decide.
What Actually Works in 2026
After enough trial and error, here's the short list worth installing on day one of any new project: RunContext7 always—the context compression genuinely reduces token overhead. GitHub MCP for code review, PR management, or repo analysis—clean API surface, reasonable overhead. Playwright MCP if you're doing browser automation at all—it integrates more cleanly than Puppeteer or Selenium and the difference matters at scale. Plus one app-specific server (Notion, Linear, or Supabase) depending on your stack. Everything else should only be added when you have a specific, measurable problem that server solves—not because it's new, not because the benchmark looks good, because your agent is failing at a specific task and this server fixes it.
The Benchmark Trap
Be suspicious of any MCP comparison that doesn't include cost-per-task. The ecosystem has developed a habit of publishing "which agent uses the most servers" leaderboards and capability comparisons measuring breadth but not efficiency. These numbers are impressive until you multiply them by actual token usage and get your monthly bill. The benchmark that matters: how much does it cost to complete a task reliably? Not how many servers connected, not how fast it runs, not which fancy new server dropped this week. Cost per task. Measured over 100 runs. With and without each server.
The Bottom Line
MCP solves a genuine problem—giving LLMs structured, reliable tool access—and the ecosystem has grown faster than anyone expected. But growth has outpaced discipline. Most teams aren't tracking token costs of their MCP setup. Most aren't auditing servers. And the benchmark conversation is all about what's possible, not what's cost-effective. The 13,000 servers are a feature and a warning. Use the protocol. Pick your servers carefully. Count the tokens. The agent running twelve MCP servers isn't better than the one running three—it's just more expensive.