Study finds median MCP server consumes 3,150 tokens per turn before any task runs

·1 views

A developer who maintains an MCP server directory sampled 60 remote servers to measure how many tokens their tool listings consume before an AI agent begins any actual work. The median server sent back enough JSON to cost roughly 3,150 tokens per turn, with the top quarter exceeding 14,000 tokens each. Notably, token cost correlated with schema verbosity rather than tool count — a 54-tool server proved two and a half times more expensive than an 84-tool one. Because tool lists are resent to the model on every conversation turn, a 20-turn chat with four average servers could burn around 252,000 tokens on capability descriptions alone. Nearly half of the sampled servers failed to respond to a basic discovery call, raising additional concerns about the reliability of MCP directory listings.

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)

WPWaaS Bets on WordPress as Infrastructure for Industry-Specific Client Software

A developer team is building WPWaaS, a platform that delivers pre-built, industry-specific software portals on top of WordPress for web agencies. The project challenges the common assumption that WordPress is unsuitable for serious application functionality, arguing it already provides reliable, familiar infrastructure. Rather than creating another standalone SaaS product, WPWaaS aims to give agencies deployable software — such as member management, bookings, and event tools — alongside the websites they already build. The team tested the concept first with a church portal, chosen because church websites quickly expose the limits of assembling unrelated WordPress plugins. The broader goal is to serve verticals like gyms and schools by treating WordPress as a foundation rather than a patchwork of disconnected components.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Embeds Ethical Philosophy Into AI Agent Instructions to Shape Its Character

A developer working on a research project called Algebraic Architecture Theory (AAT) has published an experiment in which a detailed philosophical framework was written directly into an AGENTS.md file to guide an AI agent's behavior. The framework centers on a single governing principle — speaking about software in terms that can be verified and defended — from which a set of operational vows is derived. These vows instruct the agent to prioritize rigor over output volume, distinguish between proof and hypothesis, and remain silent on claims it cannot substantiate. The experiment raises a broader question about how AI agents are evaluated, arguing that character and judgment matter alongside raw capability. The author contends that as AI agents take on more collaborative roles, the values embedded in their instructions deserve as much attention as benchmark performance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Enterprise AI Agents Gain Infrastructure, but Business Judgment Logic Remains Unsolved

The enterprise AI ecosystem is rapidly maturing in areas like agent identity, tool access, security, and runtime isolation, with Cloudflare recently open-sourcing its internal agent platform, Cloudflare OS. However, a critical gap persists between what an agent is authorized to do and what an organization should actually decide based on its own rules and context. A research paper on concurrent agent systems highlights a related risk: an agent may evaluate a decision correctly using current data, but by the time it executes, the underlying facts — such as an available budget — may have already changed. This means governance cannot rely solely on a policy check at the moment of a tool call; systems must also verify that the facts justifying a decision remain valid at the point of execution. As agents move closer to production workflows, the industry still lacks a well-defined standard for encoding and enforcing this kind of contextual business judgment.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Student Building Fake News Detector Plans to Document ML Journey on DEV

A final-year student has announced plans to share beginner-friendly tutorials on DEV Community while developing a Fake News Detection System for their capstone project. The system is being built using Python, Machine Learning, Natural Language Processing, and Django. As part of the project, the student is learning concepts such as data preprocessing, TF-IDF, model evaluation, and Django integration. The author intends to document topics that initially caused confusion, aiming to make explanations simple and practical for other learners. The initiative follows a self-described "learn, build, teach" approach targeting beginners in Python, ML, NLP, and web development.

0 comments Read more at DEV Community