Coding Agents Everywhere, Windows OS & MCP, UI In MCP

Last week, AI code editors were the theme of the week. This week, everyone's pushing a new AI coding agent.

OpenAI released Codex, a cloud-based software engineering agent. This is something of an upgrade to Codex CLI, a similar agent that runs via command line. Besides running in the cloud and skinned by a nicer UI/UX, the major differentiator between the two is Codex's ability to work on multiple tasks in parallel.

Meanwhile, Microsoft (or more specifically, GitHub) extended GitHub Copilot to serve as its own coding agent. Simply assign a task or issue to the bot in GitHub, and it will open a pull request with its solution. You can even configure MCP servers for the agent to access external systems as it works.

Google is trying to keep pace, starting to roll out global beta access to Jules, its own cloud-based coding agent. With Google I/O happening today and tomorrow (May 20 - 21), maybe we're about to see another generally available coding agent hit the market.

Our take: these cloud-based coding agents are probably nifty for small bugfixes, and maybe even building features that follow patterns already established in your codebase. And that's great: that functionality alone can save you hours per week of engineering time. The CLI tool and AI IDE alternatives are likely still the way to go for serious engineering work.

On the MCP side of things, Microsoft announced they were bringing MCP support natively into Windows OS. With these kinds of major commitments and the degree to which Microsoft has been contributing to the MCP specification, we very much appreciate having this industry giant in this open ecosystem's corner.

We at PulseMCP deployed a small design refresh on our MCP server directory: we're leaning into our concept of "server rankings" by estimated usage, and now you can see at a glance how your favorite servers stack up against each other. Love the change? Hate it? Let us know in our Discord.

What's upcoming for MCP?

→ An MCP auth debugger feature has landed in the official Inspector. It's based on the 2025-03-26 specification, so it doesn't quite serve the auth spec that is in draft, but it does align with the majority of auth-enabled MCP servers out in production right now.

→ Google's official Go team has started working in public with their design for the upcoming MCP-official Go SDK. They've chosen to implement the SDK from scratch rather than adopt an existing community solution, but are thoughtfully engaging the community for feedback and taking learnings from the community SDK's. Short term pain for Go developers while they wait, but it seems on track to be one of the more robust and well-designed SDK's in the long run.

→ The features and improvements we've written about in the past few weeks [1, 2] are all still winding their way through the PR process. OpenTelemetry, Content Capabilities, Tool Interfaces, Configuration Schema, SDK Spec Compliance, Elicitations, User Interaction. Expect some progress soon as the MCP Steering Committee has some of their first live in-person meetings upcoming at this week's MCP Developers Summit.

Featured MCP Work

MCP UI SDK by @idosal and @liadyosef

→ This is the launch we're most excited for this week: bringing UI capabilities into MCP servers. Inspired by the UI discussion we highlighted a few weeks ago, the creators of wildly popular GitMCP (Ido and Liad) have brought MCP UI to life. MCP Servers can now construct and return UI as MCP Resources, and MCP Clients can choose to render them, all with the MCP UI framework and associated libraries. This is the first step in progressing MCP past chat-focused user experiences.

Warp MCP Client

→ The well-known terminal app has launched integration with MCP servers while in Agent Mode. This setup has the potential to act much like all those agentic coding CLI tools, except in this case, the MCP integration is with the terminal itself. That has the potential to be a richer experience when troubleshooting or debugging workflows common to developers.

Desktop Commander MCP Server (#20 this week) by @wonderwhy-er

→ This one has been around for a while, but has consistently found itself near the top of MCP server usage rankings. Think of it like a well-maintained version of Filesystem, also capable of invoking terminal commands. IDE's like Cursor have a native implementation of this, but if you're using MCP outside of an AI IDE and trying to do development work, this one's a staple to include in your toolbox.

Cognee official MCP Server

→ The Cognee team was early to MCP, having launched their knowledge management server way back in December. And now they've doubled down: they dropped a major release that, among other features, has simplified data ingestion: Cognee now reads directly from your file system. We previously found the setup phase of working with Cognee to be challenging to work through, so this simplification means it's probably worth another look if you got stuck at the intimidating setup instructions in the past.

macOS System MCP Server (#59 this week) by Loopwork

→ Dubbed "iMCP", this very polished MCP server comes with a macOS app that maintains fine-grained configuration of this server. Use it to access Calendar, Contacts, Location, Maps, Messages, Reminders, and Weather. We've featured other macOS MCP servers in the past, but this one stands out for its ongoing stickiness with its userbase.

Google Workspace MCP Server (Thur, May 15; #68 this week) by @taylorwilsdon

→ While we await Google to one day release some official implementations for Google products, @taylorwilsdon released a very popular implementation this week that stretches across a suite of Google Workspace products: Gmail, Drive, Docs, and Calendar. While there are alternatives, and this implementation still requires a fairly complex API key setup, its jump into the top end of weekly usage rankings suggests it's worth a look if you've been trying alternatives.

Browse all 250+ clients we've cataloged. See the most popular servers this week and other recently released servers.

A Few Good Links

→ Kevin Indig published the first-ever UX study of Google's AI Overviews, analyzing how end-users interact with those AI results at the top of Google searches. The UX learnings translate over to how you might expect users to use chat-with-web and deep research functionalities beyond a Google search. Some highlights: (1) your brand perception matters more than ever, (2) revenue models based on eyeballs/sessions are hurting and going to continue to nosedive, (3) younger demographics tend to focus on the AI answer and then occasionally fact check it with UGC like Reddit.

→ xAI found itself at the wrong end of a censorship controversy blaming a "rogue employee" system prompt change for its Grok response bot. The change had instructed Grok to give a specific response in relation to South African genocide topics. This "rogue employee" explanation was the same excuse they used back in February for a similar incident. It makes you wonder how often this - or similar - tactics are being utilized by the xAI team and simply going unnoticed.

→ Anthropic released a Claude Code SDK, enabling builders to compose Claude Code capabilities into workflows and apps. They highlighted a use case of integrating Claude Code into GitHub Actions, which provides automated code review, PR creation, and issue triage capabilities directly in your GitHub workflow. It's an excellent building block to try shoehorning into a variety of developer workflows.

→ The Cursor team has been hard at work improving their MCP support. They launched the ability to disable individual tools, streamable HTTP support, some bugfixes, and more.

→ The big news out of VS Code this week is that they have decided to make its AI features entirely open source. They are refactoring their GitHub Copilot extension (and strong MCP support) to be core to the VS Code editor itself. This will make VS Code truly AI native and retain its entirely open source roots. It'll be quite interesting to watch how Cursor and OpenAI's recently-acquired Windsurf teams continue to differentiate against what is now a free and open source competitor from Microsoft.

→ The Dylibso team was busy on the blogging front this week, highlighted by their announcement of MCP SSO via mcp.run: they've built the ability to sign in once via their virtual MCP server, and get access to all your tools (configured and authenticated through the mcp.run platform) at once. Those tools can't include the huge mass of stdio MCP servers out there, but they do have support for remotes as well as a slew of WASM-based “servlets” available on the mcp.run platform. We also liked Dylibso's post on "a critical look on 'A Critical Look at MCP'" - largely a defense of the usage of SSE (rather than WebSockets) in the protocol.

→ We're excited to see a lot of you at the MCP Developers Summit this Friday in San Francisco. As with most MCP events in San Francisco, it is sold out, but you'll be able to access recordings of the MCP talks from Anthropic, OpenAI, Microsoft, AWS, and more (agenda). If you're looking for an MCP conference to attend that hasn't sold out yet, check out the MCP track at AI Engineer World's Fair coming in just a few weeks.

Cheers,
Mike and Tadas