SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

GitHub MCP Server Adds 62,000 Tokens of Overhead Per 20-Turn Claude Code Session

0
·1 views

MCP servers extend Claude Code's capabilities, but each registered tool injects a definition block into the context window on every session turn, creating compounding token costs. Measurements across three server setups showed that the official GitHub MCP server, with 26 tools, adds roughly 3,100 tokens per turn and approximately 62,000 tokens of overhead across a 20-turn session. For long autonomous agent loops running 2,000 turns, that overhead can reach 6.2 million tokens — costing around $18.60 at Claude Sonnet 4 input pricing before any actual work is counted. Claude Code's session logs stored in ~/.claude/projects/ allow users to inspect per-turn token usage and identify whether tool definitions are inflating costs. Developers are advised to use project-scoped MCP configs and prefer servers with fewer, focused tools to reduce unnecessary overhead.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Solo dev's critical data-loss fix sat unmerged for four days due to manual release process

A developer behind safari-mcp, a tool that lets AI agents control a user's Safari browser, discovered a serious bug on July 17 that could cause the agent to take over a tab it did not open. A fix was written and submitted the same day, passing all automated CI checks immediately, but was not merged until July 21 due to a multi-step manual release workflow. During those four days, an estimated 800 installs received the buggy version, as the project sees roughly 1,400 downloads per week. The developer also found that a prior release, v2.15.3, had shipped on July 15 with three real fixes but no changelog entry, making a contributor's work invisible to users. The incident highlights how automated checks alone are insufficient when the surrounding release process relies entirely on manual, undocumented human steps.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

WhizTools Offers 100+ Browser-Based Utilities That Process Files Locally

A developer has launched WhizTools, a free platform featuring over 100 browser-based tools for tasks such as PDF merging, image compression, file conversion, and JSON formatting. Unlike conventional online tools, WhizTools processes files directly within the user's browser, meaning no data is uploaded to remote servers. The platform covers categories including PDF, image, text, developer, security, and productivity tools, and requires no account to use. Building the client-side platform posed technical challenges such as handling large files efficiently, supporting multiple browsers, and implementing locally run OCR. Following its launch, user feedback has focused primarily on the privacy benefits, reinforcing the developer's belief that local data processing is a valued feature in itself.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

CALIO Chrome Extension Extracts Key Data from Insurance SEC Filings Instantly

A developer has launched CALIO, a Chrome extension designed to automatically extract key information from insurance SEC filings such as 10-Ks and 10-Qs. The tool was built out of frustration with the time-consuming process of manually searching through lengthy regulatory documents. Unlike AI chatbot-based tools, CALIO works directly on the SEC filing a user is currently viewing, requiring no prompts or summarization requests. The extension is available on the Chrome Web Store and currently focuses specifically on insurance-sector filings. The developer is actively seeking feedback from investors, analysts, and insurance professionals to refine what data points the tool prioritizes.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Three AI Tool Roundups for 2026 Use Incompatible Scorecards, Confusing Engineers

Three late-2025 Juejin articles each claim to answer which AI tools engineers should use in 2026, but each evaluates different categories using different scoring systems. The first ranks eight coding tools on a decimal scale, placing CodeBuddy at 9.6 and Blackbox at 7.2. The second is a consumer-chat-focused annual review recommending Gemini, ChatGPT, and Claude at roughly 140 yuan per month each. The third, a November frontend piece, rates Cursor as the top all-rounder and V0.dev as the best for UI generation using a five-star system. Despite sharing some tool names, the three roundups never cross-reference each other, leaving readers to manually reconcile incompatible frameworks before they can build a coherent 2026 AI stack.

0 comments Read more at DEV Community
GitHub MCP Server Adds 62,000 Tokens of Overhead Per 20-Turn Claude Code Session · ShortSingh