Le meilleur de l'écosystème IA et MCP, sélectionné chaque jour.
Sources
OpenAI released new audio model snapshots and expanded Custom Voices access to production voice apps. The updates improve speech quality and latency in the Realtime API and make it easier to deploy branded voices in customer-facing products. Key update for teams building voice agents or voice-first interfaces.
The MCP Transport Working Group outlines plans to evolve beyond Streamable HTTP for enterprise-scale remote deployments — addressing the stateful connection bottleneck that makes load balancing and managed services difficult. Proposes a direction that preserves backward compatibility while enabling stateless, horizontally-scalable MCP server deployments. Critical reading for teams deploying MCP at scale.
Anthropic donated MCP to the newly formed Agentic AI Foundation — a directed fund under the Linux Foundation — ensuring vendor-neutral governance for the protocol's future. MCP becomes a founding project of the foundation alongside other agentic AI standards. A major governance milestone that signals MCP's transition from an Anthropic project to a true open standard.
Anthropic surveyed 500+ technical leaders to understand how enterprises are actually deploying AI agents — 80% already report measurable ROI. The report covers common deployment patterns, key blockers (trust, evaluation, orchestration), and where agent investment is concentrated in 2026. Useful benchmark data for teams making the case for agent infrastructure internally.
MCP Specification Enhancement Proposals (SEPs) are moving from GitHub Issues to pull requests against the `seps/` directory, making proposals easier to review, version, and track. The post explains the rationale and what changes for contributors submitting new SEPs. Relevant for anyone participating in MCP spec development.
Anthropic outlines architectural patterns for building agent harnesses that remain reliable over extended runs — covering context management, checkpointing, graceful error recovery, and how to structure prompts so agents stay on track through dozens of tool calls. Drawn from experience building Claude Code and internal research agents. Pairs well with their later harness design post.
MCP marks its first anniversary with a retrospective on the year and a full breakdown of the November 2025 spec release — the most significant update since launch. Covers new primitives, transport changes, and the governance evolution that brought MCP to production at companies large and small. The canonical reference for what changed in the 2025-11-25 spec version.
OpenAI breaks down what separates good ChatGPT apps from great ones — focusing on how to design capabilities that genuinely improve conversations rather than just adding features. Covers UX principles, capability scoping, and common pitfalls from the ChatGPT Apps SDK team. Practical guidance for anyone building on the Apps SDK.
Anthropic launched advanced tool use capabilities on the Claude API — enabling agents to dynamically discover, learn, and execute tools at runtime rather than having a fixed toolset hardcoded at startup. This allows more flexible agent architectures where Claude can adapt its tool usage based on task context. A significant API capability upgrade for developers building production agents.
The original proposal for MCP Apps (SEP-1865) — introducing a standardised extension for interactive UIs in MCP, developed jointly with the MCP-UI team, OpenAI, and Anthropic. Explains the motivation, design decisions, and how the extension fits within the broader MCP extension model. The precursor post to the January 2026 official launch.
The MCP Bundle format (.mcpb) is now an official part of the MCP project — a portable package format for local MCP servers that enables one-click installation across any compatible client. Addresses the friction of manual JSON config setup and makes MCP server distribution more like installing a regular app.
.jpg)
Anthropic shares best practices for using Agent Skills to build richer, more customised frontend UIs with Claude. The post covers how skills can encode design conventions, component libraries, and project-specific patterns so Claude consistently produces on-brand output. Directly relevant to teams using Claude Code or Skills for frontend development work.
Anthropic demonstrates how combining MCP with code execution lets agents handle far more tools while consuming dramatically fewer tokens — up to 98.7% reduction in context overhead in their tests. Instead of passing tool results back through the context window, agents write code to chain tool calls and filter outputs before they reach the model. A major efficiency win for complex multi-tool agent pipelines.
An underused MCP protocol feature gets a spotlight: server instructions let you give the LLM a structured user manual for your server — explaining what tools do, when to use them, and how to combine them effectively. The post shows how well-written server instructions dramatically improve tool call accuracy without any prompt engineering on the client side.
Anthropic covers practical patterns for deploying Claude-based agents in financial services: document processing, compliance checks, customer support automation, and multi-step research workflows. The post addresses the specific trust and auditability requirements that make fintech agent deployments more complex than typical use cases. Useful for teams building regulated-industry agents.
Dagster Labs shares how they use Codex across their open-source projects to accelerate documentation — generating docs from code, translating content across formats, and measuring doc completeness automatically. A practical case study showing Codex applied to developer tooling maintenance rather than greenfield code generation.
Claude Code gained filesystem and network sandboxing, letting it run autonomously with fewer permission interruptions while remaining safe. The sandboxing isolates risky operations so Claude can proceed confidently without asking the user about every file write or network call. A key step in making agentic coding tools both safer and less annoying to use.
Anthropic launched Claude Code on the web, letting users delegate coding tasks directly from the browser without a local CLI install. Tasks run on Anthropic-managed cloud infrastructure, making it easy to tackle bug backlogs and routine fixes without setting up a local environment. Now in beta as a research preview.
OpenAI used Codex internally to build experiences, demos, and products for their DevDay 2025 event — eating their own cooking at scale. The post reveals which parts of the event infrastructure were Codex-generated and the workflows they used. Interesting as a signal for what Codex can reliably ship in a high-stakes, time-constrained environment.
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.
