AI Search Wars, IDE Mania, Official MCP Registry

In a court filing this week, Apple claimed that searches fell last month for the first time in 22 years on Safari on iOS devices. They think this is due to AI, and signalled an interest in adopting OpenAI, Perplexity, or Anthropic search services for their phones in the future.

Google denied the trend by quoting their "queries" metric. And yet, we are reminded of rumors in recent years claiming that Google has started to trade user experience in exchange for "query growth".

It sure is interesting timing for Anthropic to release a web search API this week. Combine that with ClaudeBot uncharacteristically heavily crawling the web late this week, and it seems we might have the next phase of a search-provider battle brewing.

In other AI use case news, developer IDE's had a big week. OpenAI officially closed on buying VS Code fork Windsurf for $3 billion. Cursor, another VS Code fork, has raised $900m at a $9 billion valuation. VS Code itself is keeping up in the AI race, with the team working at the bleeding edge of the MCP spec. Even Zed, the major non-VS Code-fork IDE, is establishing its relevance among AI code editors.

In the world of MCP, we finally announced our ongoing work on the official MCP metaregistry. We (PulseMCP) have been collaborating with Block, GitHub, and a slew of other contributors to start to bring this to life. When it's live, we'll have a building block in place that powers reliable MCP server discovery features across the ecosystem.

Have questions or feedback for the Pulse team? Requests for changes on pulsemcp.com? Ideas for content/collaboration? Join our Discord community.

What's Upcoming for MCP?

The registry was the big news of the week. And there's plenty of room for community involvement across a slew of open discussions and issues. In particular, we'd like to draw your attention to two:

→ If you maintain an MCP client or third party registry, we'd love to hear your use cases and feedback. Help us answer the question: will you be able to re-use the registry API's shape in a standardized way? Or should we assume that MCP Clients will not consume the registry API directly?

→ If you have an MCP server that uses some nonstandard command line scheme to start up, we want to hear from you. Help us make sure that we're catching all the edge cases as we design a standard for how MCP servers can start on a command line.

Other highlights around protocol-level progress this week:

→ The unofficial MCP Community Working Groups have started a GitHub repository explaining their mission and vision. And more specifically, they are tracking work-in-progress initiatives on this community tracker board. Keep an eye on that board to see what's getting traction for future consideration on the official track or a merged PR.

→ There is now an official Ruby SDK! Kudos to Topher Bullock and Ateş Göral from Shopify for owning and shipping it.

→ Following up on last week's work on Elicitations - the feature that enables MCP servers to request additional information from users through the client - there is now momentum to extend that further. Enter, User Interactions. While Elicitations request information through the client, User Interactions are being designed to request action outside the client. This should be very useful for concepts like payments, or collection of sensitive data.

→ @olaservo from the community moderator team has kicked off the effort to standardize a compliance schema for SDK development. With all these new MCP features flying around, it's hard to keep up: which SDK has implemented the latest, and which ones lag behind? This schema will create a source for a reliable answer to that question.

→ The handful of collaborators working on auth and MCP security have merged an initial set of security best practices, distilling down some notable attack vectors in easy-to-follow language.

Featured MCP Work

Crawl4AI RAG (Sun, May 4; #15 overall this week) MCP server by @coleam00
→ The viral launch of the week - it's gotten more use in its launch week than mainstays in the top of the rankings like Filesystem and Firecrawl. Using this server, you can (1) scrape anything, and (2) use that knowledge anywhere for RAG. Particularly useful if you are in a position where you want to "talk to a website" - and it's one that's too big to stuff entirely into your context window.

Kollectiv (Tue, May 6) official MCP server
→ From the creator of the viral Supabase server a few months ago, Kollektiv has launched to enable you to chat with and query your own documents. A dead simple UI interface lets you upload a set of files, and then you can RAG them as a group across either Kollectiv's remote SSE or Streamable MCP Server. Use it for analyzing reports, sifting through technical docs, or anywhere else you might have a handful of lengthy, private-or-curated PDFs or markdown files.

Wix (Wed, May 7) official MCP server
→ Wix announced their official MCP server as a remote, empowering their userbase to use AI-powered flows to manage website and CMS details. Managing a CMS remains a strong MCP use case (one we at Pulse love using often) - and it's becoming tablestakes for platforms to expose that functionality via remote MCP.

HTTP OAuth Example MCP server reference by NapthaAI
→ This is just a reference example, but one many developers may find useful: it's a barebones implementation of an MCP server that uses the official TypeScript SDK and implements OAuth, SSE, and Streamable HTTP. Most people building remote servers today are using a layer of abstraction, like Cloudflare's SDK. This reference helps people build without framework tie-ins, complete with a thoughtfully-written, easy-to-follow README.

Notte Browser (#21 overall this week, #2 in browser automation) official MCP server
→ Browser automation is a killer MCP use case, and Notte zoomed up the rankings for this week to wrestle the #2 spot away from Browserbase. Notte has an interesting "Notte agent" feature, which you can use to deploy an agent to complete a full end to end task on any website.

Language Server (#64 overall this week) MCP Server by @isaacphi
→ This one has been around for a few months, but picked up some newfound usage this week. And it's clever: you connect it to a local language server that enables your MCP client to use LSP-level features like `definition`, `references`, `rename_symbol`, and more. It's a more fine-grained approach that will make your coding agent significantly more efficient for certain types of coding operations.

Browserstack (Sat, May 3) official MCP Server
→ We were commenting last week how "closing the loop" on your agentic flows was one of the more powerful ways to use AI. And Browserstack is a well-known testing platform that can help you do just that while you're developing. In particular, Browserstack will shine when you're building for a range of different devices and you can say things like "run the test suite for all major mobile devices after you think you've fixed the bug".

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

A Few Good Links

→ David Soria Parra, MCP Co-Creator, was announced as the technical keynote speaker at the upcoming MCP Developers Summit. There's still some spots left, so we hope to see you grab a ticket and attend this first-ever MCP conference in San Francisco on May 23.

→ OpenAI pulled a GPT-4o upgrade that made ChatGPT "overly flattering or agreeable". While they seem to have dodged any major controversy here, it's not hard to see how this kind of miss could lead to major problems like the case of the chatbot that encouraged its teen user to commit suicide.

→ Now that MCP is coming up on 6 months old, some of the greatest hits are getting a refresh. Browserbase, which we called out as a featured server way back in December, was just completely rebuilt. We are entering the phase where companies can ditch their early proof-of-concept implementations - having shown real value to end-users - and upgrade to focusing on robustness and reliability.

→ As MCP servers mature and grow, there is a growing amount of buzz surrounding the problem of, "how do we manage large numbers of tools?" GitHub has solved it by coining a concept of toolsets, and pushing the spec on Tool Interfaces. Cloudflare solved it by releasing a separate server per API surface. Others take a search-then-notify-of-new-tools approach; or as Block is calling it "Tools like ogres… with layers".

→ Vercel is finally aboard the MCP train, releasing a guide to help developers deploy their first remote MCP server on Vercel. We agree with CEO Guillermo Rauch's take: "Good chance this becomes one of the most important frontends on the internet in coming years."

→ David Cramer, CPO of Sentry, penned a post about how MCP is not good, yet. Sentry has been one of the earliest adopters of MCP. And we actually agree with him. A lot of MCP's value remains theoretical, hidden behind layers of work that still need to be built on top of it. We're still seeing official SDK's trying to be feature-complete. Still seeing entire swaths of the spec - like Resources, Prompts, and Sampling - woefully underutilized in the ecosystem. Even yet, the best MCP servers are sticky, and adoption is growing month over month. As long as the community momentum here holds, we'll make all those MCP theories a reality together.

Cheers,
Mike and Tadas