Build1 distinct publisher2 min readUpdated
One developer's count of a five-server Claude Code setup holds up where you can check it and drifts where you cannot. The checkable part is enough to change how you connect tools.
The Engineer · Build desk

Compiled by The EngineerSomething wrong?How this is made
The only figure in that post you can rebuild yourself is the first one. The author measures a single `search_files` definition at 347 characters, about 87 tokens, and 255 tools at 87 tokens each is 22,185, which is exactly the number quoted [3][5][1]. The worked example and the total agree. That is the number to plan against.
Above it, the ledger stops closing. The extras the post itemises are server status messages at about 200 tokens each, listing headers at about 50 per server, and error handling schemas at about 100 per tool [6]. For five servers and 255 tools that is 26,750 tokens, which brings the running total to 48,935 and leaves 42,065 of the 91,000 headline unexplained [2]. The author says the counts are measured rather than estimated [12], but the itemisation supplied does not reach the headline.
The cost table is measuring something else again. At $3 per million input tokens and 20 conversations a day, the post gives $7.80 to $15.00 daily and $1,872 to $3,600 a year, excluding output [9]. Work backwards and the low end implies 130,000 input tokens per conversation and the high end 250,000 [4]. Those are conversations with content in them, not bare setups. The schema block on its own costs about 6.7 cents per conversation at that rate [7]. That is the honest line item, and it assumes the schemas are re-sent every time, which is precisely what the author's proxy claims to remove by injecting them once [10].
Then there is the awkward part. The same post counts the schemas for the same 255 tools at 22,185 tokens in one place and 2,034 in another, a factor of about 10.9 [5]. The advertised 97 percent reduction, 62 tokens against 2,034, is measured against the smaller of the two [11]. Against 22,185 it would be a larger saving, which is the odd thing about the discrepancy: it undersells the pitch while undermining the audit.
What survives is the unit of purchase. Five servers averaged 4,437 tokens of window each before anyone typed, roughly 1.3 cents per conversation per server [3]. A server is 30 to 60 tools in the author's experience [4], so the decision you are making is not "add an integration", it is "add four to five thousand tokens of permanent preamble". The post notes that the original MCP examples showed three to five tools, while real setups run to 255 [14], and its own consumer advice includes limiting how many servers you connect [13]. That advice does not require the 91,000 figure to be right. It only requires the 22,185 that checks out.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
The author connected Claude Code to five MCP servers: file system, GitHub, Postgres, Puppeteer, and a custom search tool.
The five-server setup exposed 255 MCP tools.
One sample tool definition (search_files) is 347 characters, about 87 tokens.
A typical MCP server exposes 30 to 60 tools.
255 tools amount to 22,185 tokens just for tool definitions.
Every MCP tool result is wrapped in a JSON content structure adding 47 characters of overhead, which for a 100-character result is 32 percent of tokens.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One checkable chain inside an otherwise unverified single source
Everything rests on one self-published post by the tool's author. The per-tool definition sample and its extrapolation to 255 tools and 22,185 tokens are internally consistent and independently recomputable, and the result-envelope overhead is shown in code. But the 91K headline, the daily cost band and the 97 percent compression figure each contradict the post's own numbers, there is no reproducible counting methodology, and no second source corroborates anything.
Public release only, no usage signal
A real artifact exists and is installable under Apache 2.0 with a stated test suite, and the author documents their own five-server configuration. Beyond that there is no download count, star count, external user, deployment or integration evidence in the supplied material, so adoption cannot be read above bare availability.
Headline framing overstates a real but smaller checkable finding
The framing - 91K tokens before the first question, a 97 percent reduction, thousands of dollars a year per developer, and a blanket 'all token counts are measured, not estimated' - runs well ahead of what the post substantiates. The substantiated core is narrower: schemas are a per-conversation context tax of a few thousand tokens per connected server, worth about 6.7 cents of input per conversation at the quoted price. The gap is overstatement of magnitude and of measurement rigour, not fabrication of the underlying mechanism.
Author is the vendor of the recommended fix
The post diagnoses a problem and sells the remedy: the recommended compressing proxy is the author's own project, the piece ends with install instructions, a repository link and a request for GitHub stars, and the disputed 97 percent figure is the product's headline benefit. The author does disclose non-affiliation with Anthropic and the MCP team, which is a genuine mitigating disclosure but does not address the self-interest in the proxy itself.
Low: single interested source with internal contradictions
Confidence is limited by the one-source, one-author cluster and by three separate internal inconsistencies in the central numbers. What can be held with reasonable confidence is narrow and structural - schema footprint scales with connected servers, and result envelopes add per-call overhead. The magnitude claims, the compression benefit and any real-world uptake of the tool cannot be relied on from this material.
product
A 2x LLM bill is not a bug report: token spend is an observability problem1 distinct publisher
science
OX Security says MCP command execution is a design choice, so server owners own the risk1 distinct publisher
invest
Binance gives AI agents their own subaccounts, and no loss limit1 distinct publisher
build
255 tool schemas, 91K tokens: pricing the two MCP costs nobody budgets1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 22, 2026