Workers Must Adapt, MCP User Interactions, Wielding Claude Code

It's been a relatively quiet week across the major AI labs and product teams, which probably means we're gearing up for double the fun next week.

Mainstream media didn't think it was quiet in the AI world, this week propagating Dario Amodei's soundbite from Axios and CNN interviews that "AI could wipe out half of all entry-level white-collar jobs" to great fanfare.

When you actually watch the raw footage, it's clear the take is more measured than the headlines would suggest. Yes, many of today's jobs might be automated. But humanity has proven, again and again, that it is capable of adapting to technological change. The warning here is specifically that this wave of change is coming faster than analogous changes that came before. The masses - and regulators serving them - need to pay attention, or else a good portion of them will end up spending an undue amount of time navigating unemployment.

We at PulseMCP have pushed on automating some of our own work with Claude Code paired with a $200/mo Claude Max subscription, and it is seriously impressive. When you combine this cocktail:

1. Strong automated test suites that allow Claude Code to indefinitely run in a loop until it has verifiably finished solving your problem
2. MCP servers that integrate with your external systems (e.g. deployments, observability)
3. Powered by git worktrees that allow you to easily manage 3-5 Claude Code instances in parallel

It feels like you've hired a fleet of capable software engineers that are a 10x multiplier on yourself. Much like Steve Yegge predicted. For software engineers, it's a different kind of work than what they might be used to. No more writing code, file by file. Instead, you're managing agents, feature spec by feature spec.

Software engineering is the first industry where its workers will need to adapt to a new style of work. The same for other desk jobs is coming soon.

We'll be at the AI Engineer World's Fair conference this week in San Francisco. Some talks we're particularly looking forward to: Anthropic's "What we learned from shipping remote MCP support", VS Code's "Full Spectrum MCP: Uncovering Hidden Servers and Clients Capabilities", and Pydantic's "MCP is all you need."

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?

→ A slew of official reference MCP servers have been archived, to tighten the MCP maintainers' focus on a smaller number of more manageable servers. This includes some quite popular servers, like Google Maps, Slack, PostgreSQL, and Brave Search. If you're interested in potentially picking up and maintaining something (much like is happening with the Brave Search server), we recommend you jump into the Community WG Discord and make your pitch for why you're the right person to pick it up.

→ The official MCP registry now has a list of decently well-defined GitHub Issues outlining its remaining go-live blockers. There's a fairly long list, largely seeking contributions in Go code, but also seeking input on various architectural pieces - a great entry point for anyone looking to drive some OSS contributions to MCP.

→ Elicitations have officially landed in the MCP spec. We've covered Elicitations before, but here's a new approachable writeup of how they'll work. Now, MCP servers can request additional information from end-users as they step through their tool call logic.

→ With Elicitations in place, everyone has moved on to discussing the User Interaction proposal, which is quickly ramping up to become the most-commented pull request in MCP history. Driven by Arcade.dev, folks from Google, Auth0, AWS, and more are also contributing. When this lands, we should be able to collect sensitive data (like payments, passwords) and perform step-up authorization via the end-user in the course of MCP interactions like tool calls.

→ Thanks to Shopify, there is now an official Ruby SDK! This is presumably what they've used to build out their recent Shopify Storefront MCP Server release – the MCP server setup for every Shopify store.

Featured MCP Work

Google Cloud Run (Thu, May 29) official MCP server
→ Google has entered the MCP server scene. This looks to be one of Google's first public MCP releases, helpful for deploying and managing services on GCP's Cloud Run platform. We hope this is a precursor to a slew of other servers we'd be excited about, like a remote, OAuth-friendly reimplementation of the Google Maps reference server.

Just Prompt (#26 this week) MCP server by @disler
→ Riding a wave of new popularity this week, this server allows you to process a prompt to multiple LLM providers at once (e.g. OpenAI, Anthropic, Ollama). Combine this with your development workflow to streamline those scenarios where you want "multiple opinions" on a hairy problem, perhaps one you've been stuck on. There's a cleverly named `ceo_and_board` tool that sets the "board" of models to come up with ideas, and a high-powered "ceo" model to make a decision between them.

Terraform Registry (Tues, May 27; #46 this week) official MCP server
→ Bursting onto the rankings with their launch, HashiCorp's official Terraform MCP server connects to the Terraform Registry API's, making it easy to pull in Terraform modules into your infrastructure-as-code. Much like design system/UI component libraries have become popular MCP use cases for webdevs, this server should be great for those working in the DevOps realm.

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

A Few Good Links

→ The first few presentations from the MCP Developers Summit a few weeks back are now available on YouTube. In particular, check out Shaun Smith's (@evalstate) talk on getting more out of MCP Resources. Or our talk on the official MCP registry (slides here).

→ We mentioned we've suddenly started getting a lot of mileage out of Claude Code (newly integrated with VS Code) with Claude 4 - and we're not the only ones. In particular, we love the tip to use Git worktrees to create isolated coding environments: this makes the cost of spinning up a Claude Code instance to fix a handful of small bugs basically zero, while you put most of your attention on guiding the instance working on a big feature. It's feeling hard to imagine life without the $200/mo Claude Max subscription for unlimited Claude Code use moving forward.

→ The Cline team wrote up a compelling take (twice, and in blog post format) on why Cline is not "open source Cursor" and why Cline "doesn't index your codebase." The fundamental insight: Cline as a product is designed to optimize for capabilities, no matter the cost. Compared to most tools, which are trying to thread a needle of capabilities vs. cost to support their $50/mo subscription. It's the reason there are developers willing to pay $500+/mo for API credits using Cline rather than $40/mo on a Cursor subscription. We actually think Claude Code offers a very similar dynamic here (probably to the detriment of Anthropic's financial bottom line).

→ The MCP Developers Summit has been carrying itself forward virtually with a livestreamed weekly show called The Context, which recently streamed their first two talks. Hit their Subscribe button if you want to stay in the loop when MCP industry players are on to talk about what they're working on and thinking through.

→ A thought-provoking piece out of Microsoft Entra makes the claim that OAuth needs to evolve for this incoming world of agents. "OAuth 2 works well for today's task-focused agents that act on behalf of a user. But as agents become more autonomous and capable, a new set of authorization requirements surfaces. Agents need much more granular permissions, and they need to be dynamic, easily revokable, yet auditable." This makes a lot of sense: Agents are something "in between" - not quite deterministic programs, not quite humans - and they might require a new iteration of auth infrastructure.

→ We already know it's coming, but in case there was any doubt, someone reverse-engineered ChatGPT's web app to expose a sneak peak of what their "custom connectors" powered by MCP will look like. It looks very similar to Claude's implementation: a list of predefined remote MCP servers + the option to add a "Custom" one.

→ Perplexity released a new product called Perplexity Labs. Poorly named, in our opinion, as it's reminiscent of a concept like Google Labs - experimental features not ready for general availability. But Perplexity Labs is an interesting concept: think of it like a research workspace, where you can store charts, assets, findings, related code, etc; all while continuing to do research on some topic at hand. Something like a "Cursor for research."

→ We stumbled across James Cameron's take on how copyright should work in an AI-forward world (from a Meta podcast about a month ago), and we think it's a sound idea, in theory. Humans can't violate copyright based on "inputs" and influences into their work; but they can if they produce "outputs" that bear too much resemblance to the original work. He thinks AI models should be judged by that same standard: judge them on the outputs, not the inputs. We think it's a great aspiration, but we wonder how we get there: how do you police the outputs, which are largely distributed in auth-walled silos inside chat apps where only one person ever sees the result?

Cheers,
Mike and Tadas