The key new feature in the just-released AI Shell Preview 6 is run_command_in_terminal, among several new built-in tools that allow the AI model to run commands and review outputs directly in your PowerShell terminal. Additionally, AI Shell now supports local and remote MCP servers, transforming it into a versatile AI assistant across Windows, Mac, and Linux.

Source

The ChatGPT Gmail connector was first released in June 2025 for Plus, Pro, Team, Enterprise, and Edu users. However, because of alleged data protection rules, users in the European Economic Area, Switzerland, and the UK were excluded. OpenAI is now deploying connectors for Gmail, Google Calendar, Google Drive, and GitHub across Europe for all plans—notably after Anthropic released their MCP connectors weeks ago in Europe without such concerns, suggesting OpenAI's initial data protection claims were merely pretextual or that MCP has no priority status for the AI company. The announcement mentions Google Contacts, but this connector isn't yet accessible in my ChatGPT with a Plus plan. I tested the Gmail Connector, and following GPT-5's disastrous launch, it appears OpenAI released another flop immediately afterward.

Source

Windows-MCP is an open-source MCP server that enables AI agents to control the Windows operating system, supporting tasks such as file navigation, application launching, and GUI automation. It features mouse/keyboard input tools, clipboard management, and window and app state detection. I tested Windows-MCP, one of the extensions listed in Anthropic Claude Desktop that can be installed using the free plan.

Source

💾

💾

💾

The new ChatGPT Agent Mode combines the Operator's ability to browse and interact with websites with Deep Research's ability to analyze and synthesize information from multiple sources. I tested ChatGPT's Agent Mode and was quite disappointed. Honestly, I find it surprising that OpenAI made this feature available to the public, considering that they have led the way with their chatbot so far. In my view, ChatGPT's Agent Mode is a flawed concept because it forces an AI agent to use a web browser, a relatively primitive tool designed for inferior human intelligence.

Source

I tested three MCP servers in Anthropic Claude Sonnet 4: Claude Gmail Search, Zapier Gmail MCP server, and GongRzhe’s Gmail MCP Server. I also submitted the same search prompt to Gemini in Gmail. The results varied greatly. Only two AI tools passed my test, one failed, and the most promising MCP server turned out to be useless.

Source

Anthropic has introduced two new concepts for connecting AI models to resources through MCP: desktop extensions and connectors. Extensions are available only on Claude Desktop (Windows and Mac), while connectors are supported on both the desktop and the Claude.ai web app. All desktop extensions are local MCP servers, but not all local MCP servers qualify as extensions. Conversely, all remote MCP servers are considered connectors; however, not all connectors are remote MCP servers. Although some theoretical documentation about these new concepts exists, I mainly categorized them based on how Claude Desktop organizes the different types of MCP servers.

Source

VS Code version 1.102 introduced several useful new features in GitHub Copilot. Command allow/deny lists let you specify which commands the Copilot agent can execute without asking for permission. The resubmit feature enables you to edit a previous request and resubmit the prompt. Additionally, the new VS Code version includes various updates for managing MCP servers, including a curated catalog of MCP servers.

Source

Microsoft Azure MCP Server allows AI agents like VS Code GitHub Copilot in Agent Mode to communicate with your Azure resources through natural language prompts. Azure MCP Server supports services such as Storage, Cosmos DB, Key Vault, and Logs.

Source

Recently, leading AI companies announced support for remote MCP, allowing an MCP client to connect to a remote MCP server via Streamable HTTP, which still supports Server-Sent Events (SSE). The Streamable Data Input/Output (SDIO) protocol remains essential for local integrations and remote services, which are accessible only through conventional APIs. In my opinion, the excitement surrounding new model releases has lost its significance. AI model developers seem to primarily focus on optimizing their models for AI benchmarks. However, the genuine innovations occur in the technologies surrounding AI models—most notably in AI agents and the supporting infrastructure in which remote MCP servers will play an important role.

Source

The VS Code May 2025 update (version 1.101), released a few days ago, introduces a range of new features primarily focused on GitHub Copilot. My favorite addition is the custom chat mode, which lets you create a personalized mode alongside the built-in options: Ask, Edit, and Agent. In this post, I'll also explore the new chat tool sets and the undoing edits feature. While the Simple Browser feature was included in the previous release, it deserves discussion. Furthermore, I will briefly summarize the MCP enhancements.

Source

The MCP catalog in Docker Desktop has been available for a couple of weeks now. However, MCP support only works properly in the just-released Docker Desktop 4.42. The Docker Desktop MCP server catalog significantly simplifies installing MCP servers for any MCP client. I tested it with Gordon (Docker Desktop chatbot), Anthropic Claude Desktop, and VS Code GitHub Copilot on macOS, but the process should be identical on Windows.

Source

Warp, the top terminal app with AI support, has introduced new features in its GA release, including MCP support and Refine for AI responses. The Preview release now allows image uploads and the saving of reusable AI prompts. I also found a few new features that weren’t mentioned in the Launch Log 3 post.

Source

The Gmail MCP server covered in this post lets you access Gmail from any AI client that supports MCP. I found the MCP server more capable than Google's integration with Gemini in Gmail. The AI-powered search feature proved to be helpful, but the MCP server also offers several other valuable commands. In this post, I used Claude Desktop. However, once you install the MCP server and set up the Gmail API, you can access Gmail through MCP from any compatible AI application by copying the MCP JSON. I expect that the ChatGPT app and even Windows 11 will add MCP support very soon.

Source

OpenAI recently introduced the Agents SDK, a Python-based framework that enables developers and DevOps engineers to build AI agents. The example in this post showcases a script that uses the OpenAI Agents SDK to connect an AI agent to Anthropic’s filesystem MCP server. This integration enables the AI agent to analyze the contents of a monitoring log.

Source

OpenAI's Sam Altman envisions a future where countless AI agents operate autonomously across the internet, managing tasks on behalf of users and interacting with each other to perform complex operations without human intervention. Realizing this vision necessitates a new internet protocol that supports the sophisticated communication required for AI systems. Google’s new Agent2Agent (A2A) Protocol addresses this need by providing an open standard that enables autonomous AI agents to communicate across diverse systems. This post delves into the necessity of A2A, its current supporters, distinctions from Anthropic’s Model Context Protocol (MCP), the types of data it handles, and offers a straightforward example to illustrate its application.

Source

In a remarkably short period, MCP has established itself as the standard for secure and efficient connections between AI models and external data sources and tools. Microsoft has integrated MCP into VS Code, Windows will get it, OpenAI plans to implement it for ChatGPT, and the top terminal application, Warp, has introduced MCP support in its Preview version. I installed the official Perplexity MCP server for my test and incorporated it into Warp.

Source

Perplexity is an AI search engine that answers questions by combining real-time web results with conversational, natural language responses. Installing Perplexity's official MCP server on your Mac or Windows PC allows querying Perplexity from any AI model in VS Code GitHub Copilot. Enabling your AI to search the web for the latest updates is a must-have feature for scripting or working in the terminal.

Source

I compared six frontier AI models—DeepSeek V3, Anthropic Claude Sonnet 3.7, Google Gemini 2.5 Pro, Meta Llama 4 Maverick, OpenAI GPT 4o (GPT 4.1), and Mistral Large 2—in terms of their capabilities in interacting with MCP servers. Surprisingly, the performance in terms of accuracy, token usage, and API costs are quite different..

Source

LLMs' knowledge cutoff date is a significant issue, often resulting in outdated responses. Some frontier models can browse the web, but they frequently skip this ability because they think they can handle things on their own. Bringing in an MCP server that can search the web helps the AI stay updated and assist it when struggling with a problem. Tavily is a research search engine designed to summarize online content for AI models. In this post, I will guide you through installing the Tavily MCP server on Windows. Installing on a Mac follows a similar process.

Source