MCP List-Cache Stamps Signal Intent, Not Verified Caching Behavior

·5 views

The July 2025 Model Context Protocol spec introduced list-cache stamps — ttlMs and cacheScope fields on tools/list — modeled after HTTP Cache-Control headers. However, developers are cautioning that the presence of these stamps only confirms a server made a claim, not that the claim is accurate or enforceable. A server can return well-formed cache stamps while changing its tool catalog on the very next call, and a probe checking only for field presence will still pass. The distinction matters because HTTP itself never guaranteed identical responses from Cache-Control headers — MCP imported the vocabulary without importing a verification mechanism. Engineers working with the spec are advised to test each caching clause independently, ensuring that a companion 'liar' server fails for one specific reason at a time rather than masking partial failures through combined conditions.

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)

DutchZeroHumanCompany Builds Real-Time Trust Scoring System for AI Agents

DutchZeroHumanCompany is developing a Trust Rating Agency (TRA) to assign verifiable reputation scores to AI agents operating in agentic business ecosystems. Unlike a single static number, a TRA score is a composite signal drawn from an agent's identity, credential validity, behavioral history, operator accountability, and domain-specific performance. Because agents can be updated or replaced rapidly, the system applies aggressive score decay so that recent activity carries more weight than older history, and negative events are weighted more heavily than positive ones to reflect their greater informational value. The scoring methodology is intentionally transparent and explainable, enabling operators to understand exactly why a score changed and take corrective action. This design addresses a key risk in automated marketplaces: opaque reputation filtering that could silently block capable agents and disrupt business pipelines without any visible explanation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Strava launches official MCP connector to link activity data with Claude AI

Strava has released an official Model Context Protocol (MCP) connector that allows Claude AI to access a user's Strava fitness data, including activities, distances, pace, elevation, and training trends. The integration is set up through Claude's Connectors settings on claude.ai and uses standard OAuth authorization, meaning users never share their Strava password with Claude. Access is strictly read-only and limited to the individual user's own account. Users can ask Claude natural-language questions about their training history, such as comparing weekly distances or reviewing average pace. The connector can be revoked at any time directly from Strava's account settings.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds Block Engine to run Python, JS, and Lua in a single file

A developer frustrated with writing boilerplate glue code to share data between Python and JavaScript runtimes has built an open-source tool called Block Engine. The tool allows users to write code blocks in multiple languages — currently Python, JavaScript, and Lua — within a single .blkp file, separated by language tags. An underlying C#/.NET orchestrator automatically serializes and passes variables across runtimes via an in-memory State Pipeline, eliminating the need for temporary files, local servers, or manual JSON parsing. Block Engine supports existing package ecosystems such as pip, npm, and luarocks, and can be run directly from the terminal via npx without a separate installer. The project has been released under the MIT license and is available on GitHub and npm.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Stock Music Tagger Learned His Endless Track-Skipping Was Just Procrastination

A former stock music library curator spent three years tagging tracks by genre, mood, and use case, learning to categorize songs in as few as ten seconds. After leaving the job, he noticed he would spend fifteen or more minutes skipping through music before starting personal work sessions, never actually beginning the task at hand. He initially credited this to a trained, discerning ear, but eventually recognized it as a form of self-initiated procrastination disguised as preparation. Research by Gloria Mark at UC Irvine on information workers reinforced his realization that distraction frequently originates from within rather than from external sources. He resolved the problem by removing the choice entirely, now opening every work session with the same unremarkable track that requires no evaluation.

0 comments Read more at DEV Community