200 MCP Tools Cost 24,522 Tokens Before Work

Resumo

Pixel’s announcement shares measurements of token costs for AI agents using MCP tools, from 16 tools to 500. At 200 tools, definitions consume 24,522 tokens before any work begins, while each tool adds roughly 122.6 tokens from 50 tools onward. The findings help the community understand how tool count affects context-window capacity.

DEV Community

I Gave My Agent 200 MCP Tools — It Cost 24,522 Tokens Before It Did Any Work

On Friday I measured what a handful of MCP tools cost an agent before it does any work: 16 tools from a real file server, 4,204 tokens. Fifty tools, 6,132. The numbers were small enough to shrug at, and the comments on that post kept asking the same follow-up: at what point does it stop being small? So this morning I built the same kind of server with 100 tools, then 200, then 500, and counted every token the same way as Friday, so the numbers line up exactly. The 50-tool figure reproduced to the token. What I found at the top end surprised me — not because something broke, but because nothing did, and the bill kept growing in the most boring, relentless way possible. The numbers Same ten tool shapes as Friday, repeated with numbered names after the first ten: a file reader, a writer, a directory lister, a searcher, a web search, a database query, an issue creator, a pull-request lister, a message sender, and one more. Each one has a plain-English description and a schema with two or three fields, the way real servers write them. I listed them through a real client, took the definitions exactly as a host would hand them to a model, and counted the tokens. Tools Tokens Per tool Share of 200k Share of 1M 1 101 101.0 0.05% 0.01% 10 1,208 120.8 0.60% 0.12% 50 6,132 122.6 3.07% 0.61% 100 12,262 122.6 6.13% 1.23% 200 24,522 122.6 12.26% 2.45% 500 61,302 122.6 30.65% 6.13% Read the per-tool column. From 50 tools upward it does not move: 122.6, 122.6, 122.6, 122.6. The marginal cost of one more tool is the same at tool 51 as at tool 499. There is no bulk discount and no cliff. Every tool you add costs the same as the last one, forever, in a straight line. That is the whole finding in one sentence, and it is worse news than a cliff would be — a cliff you could stay under. A straight line just keeps charging. Two hundred tools cost 24,522 tokens per message, before the model has read a single word of your actual question. On a 200k window that is 12.3% of everything the model can see, spent on the menu before the meal. Five hundred tools take 30.7% — nearly a third of the window gone before the conversation starts. What broke at the edge Nothing. I kept waiting for something to fall over, and it never did. Listing 200 tools took 14 to 37 milliseconds across two runs; 500 took 46 to 105. Registration was a few milliseconds at every size. And calling a tool, the last one in the list, the worst seat in the house, answered in under a millisecond at every size from 1 to 500. The machinery does not care. Two hundred tools is not a stress test for the protocol. It is a normal Tuesday for it. That matters because it kills the excuse I hear most often, including from myself on Friday: "the agent will handle it." The agent handles it fine. The part that does not handle it is your context window and your bill, and neither of them complains. They just get smaller and bigger, quietly, on every single message. One thing did go wrong, and it was my own probe, not the server. My first attempt at calling a tool only filled in one of its two required arguments, and the call came back with a validation error. I fixed the probe to fill every required argument and re-ran it clean. I mention it because it is the only error in the whole morning, and it was mine. What it costs in dollars Tokens are abstract until you multiply them. Sonnet 4.5 lists at $3.00 per million input tokens, and $0.30 per million for cached input. The tool definitions are the same on every message in a conversation, so they are the kind of thing that usually sits in the cache — but take both numbers, because not every setup caches them. At 200 tools, one message costs $0.0736 uncached, $0.0074 cached. A hundred messages: $7.36, or $0.74 cached. That sounds like nothing, and for one person chatting with one agent it is nothing. But this is a per-message tax, and agents do not send one message. An agent doing real work sends dozens of messages per task, and every one of them carries ...
200 MCP Tools Cost 24,522 Tokens Before Work
Powered by Readybot.io

As últimas novidades de Pixel

Incident Response: Look Beyond the Red Alert

Pixel’s DEV Community announcement explores how teams can fixate on one failing service while overlooking the real cause of an outage, such as a full …

Sach: AI Fact-Checks WhatsApp Forwards in India

Pixel’s announcement introduces Sach, a tool that checks forwarded messages in English, Hindi, Telugu, and 14 other Indian languages against official sources. It gives cited …

Java 25 Removes SecurityManager: What to Fix

Pixel’s announcement explains how Java 24’s JEP 486 permanently disabled the SecurityManager, breaking test workarounds that intercept System.exit. It recommends returning exit codes or throwing …