MCP App StoreLe meilleur de l'écosystème IA et MCP, sélectionné chaque jour.
Sources
BigCodeArena introduces a method for judging LLM code generation by executing the code end-to-end. This provides a more rigorous evaluation metric than token-based similarity.
Claude is now natively integrated with Slack, allowing teams to get help directly in channels and threads without leaving their workspace. You can also connect your Slack workspace to Claude.ai so it can search conversations and pull relevant context when needed. A significant step for enterprise teams already living in Slack.
Hugging Face introduces RTEB, a new benchmark standard for evaluating retrieval systems. This provides developers with a more robust framework to measure and improve the accuracy of AI retrieval pipelines.
Anthropic shares practical techniques for managing what goes into an agent's context window — what to include, what to summarise, what to drop, and how to structure information so agents reason more reliably. Context engineering is increasingly recognised as a core skill for agent developers, and this is one of the clearest practical guides on the topic.
A new optimization technique using depth-pruned draft models significantly accelerates Qwen3-8B agent performance on Intel Core Ultra processors. This makes high-capability AI agents more viable for local edge deployment.
The MCP team shares timelines and priorities for the next spec version — covering what features are slated, what's been deferred, and how the community can influence what gets included. Useful context for developers planning integrations or building on features that depend on upcoming spec changes.
The Swift Transformers library has reached version 1.0, providing a stable foundation for running transformer models in Swift. This marks a significant step for high-performance AI integration in Apple's ecosystem.
Hugging Face introduces Smol2Operator, a model specialized in GUI-based computer use, advancing the ability of agents to interact with visual interfaces.
OpenAI explains the design rationale behind the Responses API — how it unlocks persistent reasoning, hosted tools, and multimodal workflows that the Chat Completions API couldn't cleanly support. The post covers the architectural decisions and use cases it was designed to enable for GPT-5. The canonical reference for understanding why the Responses API exists.
Hugging Face introduces Gaia2 and the Agent Research Environment (ARE) to standardize the study and evaluation of AI agents. This provides a critical framework for measuring agentic capabilities in open science.
Anthropic published a transparent technical postmortem on three bugs that intermittently degraded Claude's responses — explaining what happened, why each took time to diagnose and fix, and what engineering changes resulted. Rare public transparency from an AI lab on production reliability issues. Useful reading for teams thinking about AI system reliability.
Hugging Face introduced LeRobotDataset v3.0, optimizing how large-scale robotics datasets are handled within the LeRobot framework to improve training efficiency.
OpenAI's engineering notes on recent Realtime API updates — covering improvements to speech-to-speech latency, turn detection, interruption handling, and session management. Targeted at developers already building with the Realtime API who want to understand the subtle behaviour changes in recent model snapshots.
Anthropic used Claude agents to help design and iterate on tool definitions for other agents — effectively dog-fooding agent tooling. The post covers what makes a tool definition good (clear name, typed params, explicit error cases) and how using agents to write tools surfaces issues that humans miss. Meta but practical.
Together AI introduces a streamlined workflow to fine-tune any model from the Hugging Face Hub. This simplifies the process of creating specialized LLMs for developers using open-source weights.
Research on training LLMs to effectively reason and execute code within Jupyter notebooks. This improves the agent's ability to handle data science and iterative coding workflows.
ModernBERT now supports multiple languages through mmBERT, bringing the efficiency and performance of the ModernBERT architecture to multilingual NLP tasks. This expansion allows for better cross-lingual representation and retrieval.
The MCP Registry launched in preview — an open catalog and API for discovering publicly available MCP servers. Provides a standardised way for clients to discover servers, check metadata, and verify compatibility without relying on community-maintained lists. The infrastructure layer that makes the MCP ecosystem more navigable as it grows.
The official PHP SDK for MCP is now generally available, built in collaboration with the PHP Foundation and Symfony. Brings idiomatic PHP support to MCP server development, opening the protocol to the large PHP developer community. Part of MCP's push to have official SDKs across all major languages.
Google has released EmbeddingGemma, a highly efficient embedding model designed for high-performance retrieval and semantic search. It provides a strong balance of accuracy and speed for RAG pipelines and vector databases.