Gemini 2.5 Image Editing, Claude Conversation Training, Grok Code

We're back after a week off last week - here to catch you up on two weeks of AI and MCP news.

Google released Gemini 2.5 Flash Image (aka nano-banana) last week, and so sets a new state of the art standard for AI image generation and editing. LMArena's community benchmark in particular gives it a resounding #1 spot in image editing. This just days after the new open source Qwen-Image-Edit model may have briefly held that slot.

Anthropic announced that they will begin training AI models on user chat transcripts. The community expressed discontent with this change, but the reality is that OpenAI and Gemini have already been doing this for some time; and all of them have opt-out mechanisms.

Perhaps more importantly, xAI's recent moves with Grok are highlighting just how valuable this kind of user usage data might be for the AI labs. A few weeks ago, xAI released a stealth model named "Sonic" for free usage. Cline reports that "the [initial] feedback honestly not great," but that after leveraging Cline's (and perhaps other sources) usage data surrounding functions like tool usage, context management, and diff editing; the official Grok Code Fast release completely flipped the narrative to "feels like an entirely different model than the Sonic I was testing".

Elon Musk piled on, making a claim that Grok Code Fast had supplanted Sonnet 4 in usage. This is misleading: it only topped the OpenRouter leaderboard because it is (temporarily) free. As we wrote about previously regarding Cursor, xAI can't claim victory here until they prove their users aren't just there for the subsidy, would they actually pay an above-COGS rate for the product?

Nevertheless, xAI has chosen to entirely subsidize weeks of free frontier model usage, and they parlayed that usage into dramatic improvements powered by real user data feedback. Combined with Anthropic's move to grab more of this same data builds a clear narrative: those contrived LLM benchmarks everyone has been chasing for the last few years are topping out. The only way to win more users is to meet users where they're actually using AI - and train (or at least tune) the models on that data.

In that realm, it's going to be hard to catch up to the kind of data ChatGPT has in the consumer world, or Anthropic's Claude Code has among developers. But the AI labs would be fools to not push to position themselves to slurp up as much of that data as they possibly can.

In the world of MCP, we're continuing to hit new all time usage highs. August brought us 20 million local MCP server downloads - up over 50% month to month from July's 13 million. And those counts have been increasing each of the past four weeks straight.

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 MCP Registry team, recently revitalized by full time effort by Adam from Anthropic (plus recent major contributions from Rado from Stacklok and Preeti from Last9) is gearing up for a Preview launch perhaps as soon as late this week. While we don't expect this Preview launch to result in instant massive adoption, it'll be a notable milestone where potential consumers and publishers can start aligning their infrastructure to integrate with the centralized work in a production-ish deployment.

→ The SEP around tool filtering / tool search / tool Groups and Tags has met some resistance from core maintainers. The discussion in Discord on the topic is still ongoing, oscillating between seeking a more comprehensive/generalizable solution, but trying to keep it as simple as possible while avoiding trapdoors. We do expect some notion of Groups to land eventually (as we've written before), given the existing precedent across the MCP ecosystem for workarounds (such as in the GitHub MCP Server's notion of "toolsets").

→ Ola from the Steering Committee has been working hard drafting a thorough blog post deep diving on the still little-used but highly useful Server.Instructions field that MCP servers can provide to an MCP client on initialization. If you're having trouble getting LLMs to use combinations of tools properly, check out the feature (and nudge your favorite MCP client builders to add support).

→ We've formally documented a process for how MCP Maintainership (and Steering Committee membership) can come to life: read here for the specific steps to take to be nominated as a Maintainer. Of course, the first step is to build a "history of merged PRs on the repositories for which Maintainership is being considered".

→ Sessions in MCP continue to be a contentious topic. This comment from Kurtis of Google builds on the state of the conversation we noted in the last edition, suggesting we do away with the idea of refactoring sessions and introduce some new concept to start peeling away the various concerns floating around. Despite this, SEP-1359, which builds on the existing notion of sessions, still seems like the leading candidate way forward, but discussion is ongoing here.

→ The existence of `structuredContent` as a possible tool call return type is continuing to sow confusion, with issues and lengthy Discord discussions revisiting the question of "should this be in the protocol?" There's arguments to both sides, and we think it's unlikely to go away: so the path forward is to write better docs and/or a blog post fleshing out the ambiguities that remain (this could be you!).

→ Auth, as thorny an MCP subject as it is, is seeing some major progress on getting thoughtful, thorough documentation into place. This monster authorization guide kicked off by Core Maintainer Den (Microsoft) has been through a number of rounds of feedback, and should land soon. In the meantime, Core Maintainer Paul (Anthropic) pieced together a set of MCP OAuth Examples for those evaluating auth options.

Featured MCP Work

Obot MCP Gateway by Acorn Labs
→ Another week, another MCP gateway launch. We think the team behind Obot is a good one to back: Acorn Labs is the company driving the flagship MCP Dev Summit conference through substantial financial and manpower contributions for the benefit of the MCP open source community. And we like their take on MCP gateways: they're not trying to pitch nice-to-have, solo developer features like "manage all your MCP servers in one place". Their explicitly stated target audience: IT admins at enterprises. Those are the people who will benefit from controls on greenlighting pre-vetted MCP server usage across enterprise teams, unifying configurations, managing audit logs, and more. And it's open source!

MCP Client Capabilities Package by Apify
→ The community-managed MCP Clients Table has made a valiant effort keeping up to date with the degree to which each MCP client app has adopted the MCP specification, but it's only a manual reference. This new package from Apify turns that kind of client capabilities data into a tool MCP server implementors can use to properly flag certain features or gracefully downgrade to methods that a particular MCP client could support. In theory, this should be possible to do with the protocol itself, but in practice, the protocol has a lot of maturing to do before we have a proper solution in place that is widely supported across servers and clients. The catch with Apify's package: it needs community-submitted data to succeed. If you'd find this kind of tool useful, we encourage you to use it and open some PR's to help Apify make it broadly useful and accessible.

Google Sheets (#53 this week) MCP Server by @xing5
→ Not a new server, but regularly maintained and showing consistent slow growth to finally break into the upper echelon of usage metrics. As usual with third party Google products, authentication is a bit of a pain, but the included README is very thorough - and once you're set up, you're free to wrangle those Google Sheets from the comfort of your favorite MCP client.

Widget MCP Server by Ref
→ This is just a demo server, but MCP-UI - the idea of embedding UI widgets in MCP-powered LLM conversations - continues to be a hot topic, so we'd be remiss to not plug another showcase of it here. This demo showcases some traditional widgets you might be used to seeing in Google Searches (timers, conversions, etc) - expect more and more MCP servers to start offering UI capabilities as major clients (like Goose) stand up support. And maybe some MCP servers that wouldn't exist altogether without a concept like MCP-UI…

Browse all 300+ clients we've cataloged. See the most popular servers this week and other recently released servers.

A Few Good Links

→ CDN providers seem to be following Cloudflare's lead in investing in anti-AI-crawling measures. Fastly now offers similar features. Akamai too. Customers are actually using the features to great effect. Browserbase is partnering with Cloudflare to push the envelope for properly identifying bots and the various different forms they take. We think this is the right way forward: force the industry to get fine-grained so that every website has complete control over who can and cannot programmatically access it. In many cases, an easy solution might be to offer an MCP server, and otherwise leave your website for human-only access, no agents allowed. Some companies will overstep, prioritizing human ad placements over agent-friendly (perhaps still ad-laden!) UX: those companies will lose the free market battle over time. We're excited to see the open web continue to flourish with these efforts to align everyone's incentives.

→ In Google Search news, Google claims that "people really prefer and are more likely to click links that are embedded within AI Mode responses." They're using this insight as justification for building out the AI Mode UX to surface more links and nudge users to click them. Good news for publishers and website owners: if this is truly a reliable behavior pattern of AI Mode users, then whether it's Google, or OpenAI, or Claude driving traffic to websites, they should all converge on the same insight: end-users still want to visit websites. Of course, there's likely nuance to the general idea that clicks are generally desirable, but it does offer a shred of hope that there may be a path to LLM-answer-engine to content creator incentive alignment.

→ In other Google news, a federal judge ruled that Google does not have to sell off Google Chrome, putting an end to that recent unsolicited bid to purchase from Perplexity. This is a big win for Google as it reverses a previous court decision that it would have to do as part of its antitrust case that declared Google a search monopoly. However, the decision still maintained that Google "will have to make certain search index and user-interaction data available to certain competitors". This is major news for prospective rivals like OpenAI, Claude, and Perplexity in AI-powered search: as we saw in the land-grab of user data at the beginning of this edition, getting access to Google's storied search-related user data could poke a massive hole in Google's monopolistic search moat.

→ A report that "95% of generative AI pilots are failing" hit the mainstream media a couple weeks back. But what we find interesting is the implication that 5% are succeeding. We've been witness to a large skill gap variance across AI practitioners, such that it is categorically unsurprising to us to hear that only 5% of the most forward thinking have properly threaded the needle from new capabilities through to business value in the relatively short time that AI capabilities have been production-ready. A great example of a company that has truly figured it out: Intercom. They've gone from near-death at $1m ARR to $12m ARR in 12 months, boasting impressive ongoing growth to boot. All on the back of Fin, their flagship AI agent.

→ Better late than never: Elon is delivering on his promise to open source old models of Grok as newer ones get released: Grok 2.5 is now open source. He originally promised they would open source the older models as soon as they release new ones, and instead they seem to be taking a tactic of open sourcing two steps behind. This trend inspires confidence in the theory we've been pitching: there's little harm in relying on expensive models to get your work done today, because your API fees for the same work will go to zero in a mere year or two.

→ Nonetheless, the reality is that we're probably never going to be satisfied with using just last year's frontier models, no matter how cheap they get. Aaron Levie puts it well: "the general cycle [that we spend more and more on AI inference] will go on essentially forever, because we will just keep raising the bar of what we do with AI." Costs per unit of work will go down - down to zero for some particular units of work. We're just going to keep getting more and more work done.

→ Burnout is a concept common to builders and leaders, and the breakneck pace of the AI world we're in is likely particularly susceptible to falling prey to it. But we thought this writeup from the CEO of Boom was a worthy read, especially if you may have tussled with burnout lately. He makes the claim, "Burnout is not what it presents: it's not about working too hard for too long, burnout is about working in the face of a goal that seems too far out, too unattainable, too abstract". Sometimes, the antidote to burnout may be simpler than taking your foot off the gas: maybe it'd be more productive to reframe your goals or pivot your strategy.

→ If you haven't bought your tickets yet, MCP's flagship event, the MCP Developers Summit, is upcoming in London on October 2. If you loved the first one in San Francisco in May, you won't want to miss the European version. Sign up for tickets, or check out the early speaker list.

→ If you're in San Francisco, come see yours truly (Tadas) on the panel at Speakeasy's MCP After Hours event next Tuesday the 9th!

Cheers,
Mike and Tadas