MCP benchmark: ten popular servers inject 111K tokens before your first message

What was measured

On 23 August 2026 a developer publishing as "mcptokensaver" on dev.to posted a three-part benchmark measuring how many tokens popular Model Context Protocol (MCP) servers push into an agent's context window before the user has asked a single question. The method was straightforward: install the ten most popular servers from the official registry, connect over stdio, call tools/list on each, and count the results with tiktoken's cl100k_base encoding. Tool-result overhead was measured by calling tools/call twenty times per server.

The headline figure is 111,713 tokens of JSON schemas across the ten servers, with Google Drive alone accounting for 47,293 tokens for its 31 tools. One of the posts puts the total above 200,000 tokens once server status messages, headers and error schemas are included.

Server	Tools	Tokens
Google Drive	31	47,293
Slack	22	14,672
Notion	24	13,780
GitHub	28	12,440
Postgres	19	8,231
Puppeteer	15	5,890
Filesystem	11	3,847
Memory	9	2,567
Brave Search	8	2,103
Sequential Thinking	3	890

The posts are not fully consistent with each other: one claims 847 tools in total, but the per-server tables in all three sum to 170.

Where the tokens go

According to the author's breakdown, 42 percent of the schema cost comes from inputSchema properties, 35 percent from tool names and descriptions, 15 percent from nested type definitions, 3 percent from required-field arrays and 5 percent from server metadata. In other words, the largest cost is not explanatory prose but parameter structure: every argument needs a type, a description, sometimes enums and nested objects.

A second inefficiency is wrapping. Each MCP tool result arrives inside a content envelope, and in one measured example the payload was 38 characters while the wrapper was 47. Across twenty calls the author estimates roughly a third of result tokens are pure envelope.

What it costs

At Claude 3.5 Sonnet pricing of $3 per million input tokens, the ten-server setup costs about $0.34 per conversation in schemas alone, or about $0.54 once twenty tool calls are included. Annualised at twenty conversations per working day, the posts arrive at different totals — $2,376 in one and $2,592 in another — so somewhere in the mid-two-thousands of dollars per year.

The context claim is starker. Roughly 111K schema tokens plus Claude Code's roughly 8K system prompt leaves about 81K of a 200K window before the first user message. The author reports conversations degrading around message 15 in that setup, and lasting to roughly 45 messages after cutting back to three servers.

The wider debate

The benchmark is framed against public criticism the author cites but which cannot be verified here: Y Combinator CEO Garry Tan writing "MCP sucks honestly" on X, Perplexity CTO Denis Yarats saying the company is replacing MCP with REST APIs and CLIs internally, and an Anthropic engineering post reportedly showing complex MCP workflows consuming over 150,000 tokens, with a 98.7 percent reduction achieved by moving tool definitions out of context and into runtime code. Cloudflare is invoked too, with a 2,500-endpoint API that allegedly needed 1.17M tokens of schema compressed to about 1K by exposing just two functions, search and execute.

The fix on offer, and the caveats

All three posts promote the author's own tool, mcptoon, an Apache-2.0 licensed CLI proxy that caches schemas across conversations, strips result wrapping and compresses tool definitions into a compact "TOON format" — claiming 111,713 tokens down to 3,247, and per-conversation cost from $0.54 down to $0.14.

The cheaper advice in the same posts needs no tool at all: audit your config and disconnect what you do not use. The author's own audit found three of ten servers were used zero times per week, including a 47K-token Google Drive server used roughly once a month.

The caveats matter. This is a single author's self-reported data published alongside promotion of their own product, the three posts disagree with each other on tool counts and annual figures, and tiktoken's cl100k_base is OpenAI's encoding rather than Anthropic's tokenizer, so exact counts on Claude will differ. The numbers are best treated as indicative until independently replicated.

Why it matters

MCP is becoming default plumbing for AI agents, and the structural problem the benchmark points at is real regardless of the precise figures: every connected server's full schema is loaded into context at startup, so token cost scales with the number of tools exposed, whether or not they are ever called. That taxes both budgets and reasoning quality, because schemas compete with the actual conversation for window space. For server authors the practical guidance is to keep descriptions short, flatten schemas and avoid exposing unused tools; for users, connecting fewer servers is free and immediate. The fact that Anthropic and Cloudflare are already architecting around the problem suggests the pressure is felt well beyond one developer's invoice.