OpenAI Adopts MCP, Gemini Pro 2.5 Codes, Studio Ghibli Images Abound

Studio Ghibli style images took the internet by the storm this week on the heels of OpenAI's release of native image generation in their 4o model. Only two weeks after impressive image capabilities from Gemini 2.0, OpenAI's 4o leapfrogged those capabilities. Generate and modify all kinds of images with ease (just beware of the copyright battles that are sure to come).

Gemini released 2.5 Pro Experimental, a "fundamentally thinking" model. And the reviews are in: it and its 1m token context window is excellent at coding. The bleeding edge pattern appears to be to use 2.5 Pro for planning, and Sonnet 3.7 for executing. Though some seem keen to want to switch to 2.5 Pro altogether.

All that said, 2.5 Pro hasn't really reached the masses yet, as its usage has been plagued by major capacity and rate limiting issues. Speed and reliability of a model is as important as any metric for assisting developers, so we'll see if they can turn that ship around quickly enough.

But the big news this week for our world: OpenAI has officially announced adoption of MCP, combined with integrating it into their Agents SDK. This means MCP is certainly here to stay. Kudos to OpenAI for being a step ahead of Gemini, Grok, and the rest of the AI field on the adoption curve.

Did we miss something this week? Notice something interesting you want to make sure we cover next time? Shoot us a note, we love feedback.

What's upcoming for MCP?

The 3-26 release of the MCP specification has landed. The major updates:

→ Big changes we've covered before. Authorization via OAuth. Streamable HTTP to make servers stateless.

→ ToolAnnotations snuck into this release as well. As per notable demand from the community, servers can now label individual tools with tags like "destructive", "idempotent", "openWorld", and "readOnly". Clients are meant to treat these annotations as hints (i.e. they should not trust untrusted servers to self-report properly), but it should unlock some great efficiencies with respect to dodging the alarm fatigue you get when having to repeatedly click "Allow Tool Call".

While a great step forward for the specification, don't forget that it'll still be a while for end-users to be able to take advantage of all this. First, the SDK's (e.g. TypeScript, Python) need to implement the new auth, streamable, and annotation features. Second, framework-layer providers (e.g. Cloudflare, Mastra) need to update their implementations to support the new features. Third, server developers need to update their server implementations accordingly. And fourth, client developers need to update their client implementations accordingly.

To that end: it's worth pointing out that because the official SDK's do not yet support the new Streamable HTTP specification, any product claiming to work with "remote MCP" is likely operating on the now-deprecated SSE specification. In some cases, that may be fine to adopt or integrate, but be wary of the distinction and backwards-compatibility story if you start building on top of it.

The MCP core team has also begun laying out its official roadmap for the next 6 months. Some potential highlights: reference client implementations (going beyond Claude Desktop), official MCP registry, and video modalities.

Featured MCP Clients

VSCode by Microsoft
→ Microsoft is gearing up to "ship MCP support in VSCode" next week. Expect easy access to install, manage, and use MCP servers, all natively available in VSCode. The VSCode team has been working on a slew of MCP-related features for weeks now, and seems ready to release it out into the wild.

Workers AI LLM Playground by Cloudflare
→ Simple but reliable and web-native: use this playground to test out the first remote MCP servers coming online. It uses open-weight models like Llama, so the tool calls won't be as robust as something like using Sonnet 3.7, but it's great for quickly testing whether the SSE URL you've found for your favorite MCP server is working properly.

Fast Agent (Updates) by @evalstate
→ fast-agent is the only MCP client in existence that supports all 5 of MCP's major features: resources, prompts, tools, sampling, and roots. It's not surprising; @evalstate has continued to develop at the bleeding edge of the specification and help push it forward. Although very few servers can match this client support, you know you can lean on fast-agent to be your first test case as you develop servers with these capabilities to match.

Featured MCP Servers

Ghidra (Tues, March 25) by @LaurieWired
→ Ghidra is a software reverse engineering framework. If you have a program that's compiled down or obfuscated, you can use it to reverse-engineer its source code. Which makes for a great LLM use case: skip the tedious labor of reverse engineering everything yourself and let Claude do it for you. Laurie got a lot of love on HackerNews for this server.

Wireshark (Fri, March 28) by @0xKoda
→ If you're no network engineer and you've ever tried to debug some networking issue by running packet captures and other over-the-wire data, you know how much of a rabbit hole it can be. No longer: use this server to let an LLM leveraging Wireshark do that analysis for you.

Pipedream (Fri, March 28) official implementation
→ Pipedream, a company that has been building towards the promise of "connecting applications to all the services in your stack", is going all-in on MCP. They've bridged their platform to the MCP model, exposing 2,500+ auth-capable API's as MCP server tools. It's a model similar to that of Zapier or Composio, where you have a centralized platform from which to fan out your MCP usage.

R Econometrics (Thurs, March 27) by @goji+
→ Big, complex economic datasets require tools like the R software environment to sift through. Using this MCP server, economists can leapfrog the quirks of R and perform analyses like regressions, tease out biases in datasets, and more with natural language.

PlayCanvas (Sun, March 23) official implementation
→ This idea of building MCP servers to chat with external services over WebSocket plugins is getting great traction. Last week, it was a novel approach to a Figma server. This equivalent take on PlayCanvas offers a path for game developers to control their WebGL authoring environments with natural language.

Arduino (Wed, March 26) by @vishalmysore
→ If you've been having any doubts as to how AI-powered robots might be dangerous to humans in the not-so-distant future, now you can run some tests at home with Arduino hardware. Hook up those Arduino motors to Claude and see what a runaway AI might be able to do before you can reach for the "off" switch.

A Few Good Links

→ Anthropic released notable research this week on tracing the thoughts of large language models. Check out this well-crafted video explainer. It starts to unravel how these LLM's truly "think" - and how potentially naive attempts at "chain of thought" analysis actually serve to surface "plausible-sounding steps" that are misrepresentative of how the LLM is actually making decisions. And they're objectively smarter than "next token" predictors: they're capable of thinking ahead.

→ mcp.run wrote up a great blog post distilling down why MCP is such a powerful and needed abstraction layer. This is not another one of those "what is MCP" puff pieces. Its core insight and takeaway: "Focus on intent over implementation: Design your tools around what they accomplish, not how." It's a different way of thinking about designing MCP servers to not be naive (and ultimately hardly helpful) 1:1 mappings to REST endpoints.

→ This writeup by Annie Vella strikes at the core of the fear many software engineers might be having: an identity crisis over the prospect of AI taking over the most-loved parts of their coding craft. Annie does a good job acknowledging that the engineering profession isn't going away - the managerial and orchestration aspect of it has become only more needed and powerful. But the problem remains is that writing the code is what so many engineers love and consider their craft: and that is certainly at risk of losing its economic value.

→ In an AI-forward world where unique, original data is increasingly worth its weight in gold, genetic testing company 23andMe is certainly sitting on a mountain of it. That genetic data is the property of its individual users… except now 23andMe is heading towards bankruptcy. And who knows what bad actor may end up being the highest bidder. If you've ever used 23andMe, this is a PSA to go to 23andme.com → profile icon → Settings → 23andMe Data → Delete.

Cheers,
Mike, Tadas, and Ravina