MCP Protocol Injects 111,000 Tokens Before You Type a Word, Analysis Shows

·2 views

A developer analysis found that connecting 10 popular MCP servers to an AI model injects over 111,000 tokens into the context window before a single user message is sent. At Claude 3.5 Sonnet pricing, this overhead alone can cost users roughly $1,764 per year across a typical workday usage pattern. The Model Context Protocol, backed by Anthropic, OpenAI, and Google, has amassed 97 million monthly downloads, but critics argue its design assumes context is free — an assumption that breaks down in complex, multi-turn agent workflows. Anthropic's own engineers demonstrated token consumption could be cut by 98.7% by abandoning schema injection in favour of runtime code execution, while Cloudflare compressed a 1.17-million-token API surface down to just 1,000 tokens using a two-function wrapper. These findings have prompted renewed debate about whether MCP's architecture is fundamentally flawed despite its widespread adoption.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

JSONLinter.io Offers 42 Privacy-First JSON Tools With No Server Data Uploads

A developer has launched JSONLinter.io, a browser-based JSON toolkit that processes all data locally without sending it to any server. The platform includes 42 tools spanning six categories, such as validation, formatting, data conversion, code generation, schema tools, and encoding utilities. It supports conversion to and from formats including CSV, YAML, XML, and SQL, and can generate typed code in languages like TypeScript, Python, Go, and Rust. The tool is built using React 19, TypeScript, Vite, and Tailwind v4, and runs entirely client-side without requiring any signup or installation. An optional AI assistant is available on the validator page, and the project is accessible at jsonlinter.io.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Fix Next.js Hydration Mismatch Error in App Router

A common Next.js error occurs when server-rendered HTML does not match the React component tree built during client-side hydration, breaking user experience and causing unpredictable behavior. The mismatch is typically triggered by dynamic values like Date(), Math.random(), localStorage, or browser APIs used directly during rendering. The recommended fix is to initialize component state with a safe, static placeholder and update it only inside a useEffect hook, ensuring both server and client produce identical output on the first render. Developers should also add a format-detection meta tag in the root layout to prevent iOS from silently injecting nodes around dates and phone numbers. Additional causes include HTML minification by CDNs like Cloudflare or Vercel Edge, and browser extensions that modify the DOM, both of which should be investigated when the error appears only in production or development respectively.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

MCP Server Schema Bloat Can Consume 60% of Claude Code Context Before Work Begins

A developer discovered that running 10 MCP servers in Claude Code was injecting over 111,000 tokens of schema data at the start of every conversation, consuming roughly 60% of the available context window before any actual work began. Measuring each server individually revealed that three were used zero times per week and two more only once or twice, yet all loaded their full schemas every session. The developer estimated this silent overhead cost approximately $2,500 per year based on per-conversation token pricing. Reducing active MCP servers from 10 to 3 cut context overhead by 97%, extending usable conversation length from around 15 messages to 45. An open-source tool called mcptoon was also highlighted as a way to audit server token costs and compress remaining schemas to reduce waste further.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

MCP Servers Can Consume Over 111,000 Tokens Before You Ask a Single Question

A developer audited 10 commonly used MCP servers and found they collectively inject over 111,700 tokens into an AI agent's context window before any user query is processed. The Google Drive MCP server alone accounted for 47,293 tokens — roughly 5% of the entire works of Shakespeare — just to list files. At Claude 3.5 Sonnet pricing, running 10 MCP servers across a typical workday could cost up to $2,376 per year. The token bloat also degrades agent performance: with heavy MCP usage, over half the context window is consumed by tool schemas, reducing reasoning capacity and conversation longevity. The author found that several of their own MCP servers went unused weekly and had direct CLI alternatives, prompting a shift away from MCP-based tooling.

0 comments Read more at DEV Community