Claude Agent Skills, Ads in Amp, Agentic Engineering Efficacy Debate

Anthropic released Agent Skills this past week. The idea here is to propose a new level of abstraction, "skills", that are defined in freeform text (merely a name and a description) and attached to a set of supplementary resource files (usually markdown or scripts). Using that structure, any agent capable of reading files and executing arbitrary code can be "taught" how to accomplish some class of task in a specific way.

Particularly compelling is Anthropic's guidance on building Skills. Start with evaluating how your agent performs. Wherever it gets stuck, add a Skill. This comes off as an excellent, flexible way to evolve a base agentic loop like Claude Code into something that meshes perfectly with your bespoke codebase and your (or your team's) best practices and optimized workflows.

As the concept is being lauded as potentially a "bigger deal than MCP," that notion has been misleadingly parroted as a "replacement of MCP". They serve two different roles: MCP is largely for hooking into external tools or agents on terms set by the external service. Skills sound great for flexible, highly customizable flows specific to the end-user. Use skills when you want to codify logic specific to how you work. Use MCP servers when you want to give agents access to a service.

Admittedly, Skills do encroach on a class of MCP servers that did little more than inject a logic sequence, or servers that were very bespoke to specific use cases. This may be the death of popular servers like Sequential Thinking or Time. They even align well with the concepts we laid out a week ago when describing a future with Agentic MCP Configuration: our theory there remains sound, but the proof of concept implementation would do better to use Skills (not a custom MCP server) as its top layer. MCP remains critical to the lower layers.

The ads we've been expecting in the AI space for the last few months are slowly but surely arriving. First up: Amp, the coding agent. They've launched a Free tier that is ad-supported. You get frontier models at no cost, in exchange for seeing a developer-target ad in an unobtrusively-styled corner of your Amp screen.

It's somewhat surprising to see a coding agent be the first notable AI product to take this step, given that developers are a target persona well-known to be particularly allergic to ads. But developers don't like to pay for software either, so it's a smart bet on Amp's part, as a counter position to the coding agents that are running developers hundreds of dollars per month.

Meta appears to be not far behind, announcing a few weeks ago that they would start using users' chats with Meta AI in ad targeting efforts. Surely, ChatGPT must be making the same plans. If mainstream media is worried about being in an AI bubble now, just wait until ChatGPT, Gemini, and Claude start collecting and reporting ad revenue: then the valuation frenzy can truly begin.

Agentic engineering practices have been a major topic of recent discussion - and the field seems split on whether great engineering means short spurts of AI assistance on a tight leash (Andrej Karpathy, Quinn Slack), or if the real unlock is wielding long-running asynchronous agents - oftentimes many in parallel (Gergely Orosz, Simon Willison).

We think they're both right. Software engineering is a vast field. Karpathy is probably someone who doesn't spend a whole lot of time building CRUD apps. But the long tail of CRUD apps is what brings the lion's share of software engineering's economic value to life. The frontier agents of today are certainly capable of autonomously building large swaths of those CRUD apps with minimal intervention - and beyond that it's unsurprising that the level of original thought that goes into more highly technical products still requires more hand-holding and an eye towards human-driven (rather than human-reviewed) development.

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?

→ Core Maintainer Den Delimarsky and a boatload of community contributors have finished a two month marathon to land a full MCP authorization guide into the official documentation. This is now the canonical place to go if you're looking to ramp on or implement auth in MCP.

→ In the same breath, the core maintainers have approved SEP-991 to bring in CIMD (Client ID Metadata Document) as a robust alternative to the much-maligned DCR (Dynamic Client Registration) approach to OAuth in MCP. This came shortly after CIMD landed in the OAuth spec itself. That means that authorization guide is going to need a new CIMD section soon - but hopefully this step puts most of the thrashing around auth and MCP to bed, for good. Just ecosystem adoption left to go. VS Code is leading the charge as usual on the client side: VS Code Insiders already supports it.

→ The MCP Registry's initial commitment to avoiding breaking changes in the API is imminent. This is not yet a General Availability launch - but you can consider it a release candidate launch. We recognize all the iterations have made it difficult for sub-registries to properly integrate, and we hope this milestone of stability will lead to the first meaningful integrations of the Registry in the wild. If all goes well, we'll be tracking to a General Availability launch with no further breaking changes.

→ Core Maintainers David from Anthropic and Nick from OpenAI have teamed up to start a SEP around Server Cards, and hosting them at a .well-known URI. We expect this SEP to converge with the Registry's notion of server.json that we have been evangelizing for the last few weeks.

Featured MCP Work

Gemini CLI Extensions and Claude Code Plugins

→ Two sides of the same coin, both of these coding agents launched very similar features within a day of each other. They each enable a shareable way to extend their respective coding agent's functionality, with separate takes on what data is needed (Gemini: context files; Claude Code: skills, hooks, agents) as well as a couple common threads: commands, and, of course, MCP servers. Paired with the community's push to get the MCP Registry in production-ready shape, we hope extensions like these are soon much easier to wrangle; the "MCP servers" section should be merely a name declaration away from installation.

OpenAI Apps SDK by Fractal

→ Many developers are digging into building OpenAI Apps, and Fractal is making that process much simpler with a ready-to-use open source SDK, largely focused on making it easy to build these Apps with React. Get started by wrapping your head around the technical weeds in this writeup by Alpic and MCPJam, and then jump into Fractal's quickstart to get an App up and running in no time.

Ahrefs Official MCP Server

→ Ahrefs, a leading SEO research software suite, has a thoughtfully designed MCP server for which Glen Allsop recently published a slew of SEO research use cases. We spent a portion of our careers working in SEO, and it's striking how those use cases align with major chunks of SEO deliverables that form the bread and butter of many SEO agencies out there. Of course, the analyses and data wrangling the MCP server can do is hardly valuable without a domain expert guiding the conclusions, but that "report on SEO strategies of top competitors" is now just a prompt away: no Ahrefs SEO consultant expert required to slog through hours of hard-to-learn Ahrefs dashboards. If an MCP server can achieve this in the SEO industry, the same is inevitable in so many more.

Sora MCP Server (#85 this week) by @Doriandarko

→ Although it presents as a simple API wrapper on a viral new product (the Sora 2 API), what's interesting is how popular the server is anyway. MCP is no longer a novelty that you might expect users use just to experiment with the concept of an MCP server: the continued usage of tools like this goes to show how sticky the MCP-install-and-try workflow can become. Indeed, we've regularly found ourselves reaching for basic wrapper MCP servers simply because the familiar interface of our go-to MCP client (Claude Code or Claude Desktop) is an appealing way to interface with some otherwise-unknown service.

Browse all 400+ clients we've cataloged. See the most popular servers this week and other recently released servers.

A Few Good Links

→ Besides Amp's headline news that they're funding Amp Free by running ads, they subtly include a mention in their announcement that, in addition to ads, Amp Free is funded by the "sharing of training data." We've written before about how valuable this data is to frontier model builders - and we wonder if that revenue stream might actually form a fairly significant portion of this Free tier's funding. Of course, ads have more upside: that market will only grow, while training data needs may falter or hit a ceiling over time.

→ The Wikipedia team penned a post lamenting that generative AI has caused an 8% drop in its human traffic. The theory is, of course, that generative AI experiences are answering end-user's questions - and either not attributing the answers, or including mere citation footnotes that users are not clicking. This is reminiscent of Reddit's analogous problem, in response to which they are leveraging their big AI data contracts to explore UX that will encourage end-users to contribute back to Reddit. This is clearly a problem that is going to plague any user-generated content-reliant business - and it's probably going to get worse before it gets better. Perhaps companies that happen to be MCP-inclined would do well to push the MCP ecosystem to invest more in elicitations: they at least provide the promise of a UX that funnels data from end-users back to service providers.

→ Anthropic released Claude Code for Web - bringing the Claude Code experience to an elegant web interface. We think it's likely a helpful iterative improvement on how you might've been using Claude Code in GitHub, but might not be a great stand-in for much local development if you've already gotten familiar with something like a custom git worktree-powered setup. Like they said in their announcement post, the primary use cases will likely be lower touch "answering questions about how projects work," "bugfixes and routine, well-defined tasks," and the like.

→ The Cognition team released SWE-grep, a model specialized for highly parallelized context retrieval. Their goal was to enhance the "agentic search" flow in a codebase context, and they succeeded: it's an order of magnitude faster with no loss in performance (though the OpenCode team was quick to claim they've got something better coming). We love iterative innovations like this. It's a great example of how a team can take a slice of repeatable work (agentic codebase search), prioritize lowering cost and increasing speed, and then drop it into existing workflows. With new patterns like Agent Skills, this kind of improvement should be getting particularly easy to drop-in boost performance (and lower costs) for a team using AI heavily.

→ Mitchell Hashimoto, creator of Ghostty, the popular terminal emulator, wrote up a thorough play-by-play of using a coding agent (Amp) to build a fairly complex feature for Ghostty. It demonstrates how using AI was a clear productivity boost for him, but wasn't as simple or straightforward as much of the social media discourse might lead you to believe on the subject of vibe coding. Reading this post is a helpful goalpost around which to orient any debate around agentic engineering: we think Mitchell is squarely middle of the spectrum insofar as his (un)reliance on AI - and it's around here that most conversations around "should I use AI agents for coding" should start.

→ Cloudflare's post on "Code Mode" for MCP about a month back generated a significant amount of anti-MCP sentiment: if code is the right abstraction, why bother with MCP's complexities? What those people missed is that Code Mode is not a critique of MCP: it's building on top of MCP. Adam Azzam put it well: "MCP standardizes how a client discovers what tools exist, fetches schemas, authorizes access … Once you have capabilities, you choose how to present them to the model, [of which generating TypeScript code, as Cloudflare's post suggests, is one good option]." That good option falters in many scenarios, despite performing well in others, and so only suits a small subset of AI <> MCP possibilities.

→ Stripe put out a stats teaser that their MCP server has been putting up impressive growth numbers: 32% week-over-week growth in number of users, and 24% week-over-week in tool calls. And they're not starting from a low baseline - Stripe's MCP server is in the top 30 most-downloaded of all time, and that's not even accounting for their remote variant. Stripe happens to be well-positioned with a developer target audience (and most MCP users today happen to be developers), but we think these kinds of numbers may be a good benchmark for what an MCP server with good product-market-fit might look like. We'd expect to see complex-admin-dashboard-distilling servers like Ahrefs, or something like product analytics tool Amplitude, having similar traction.

→ In case you missed it last week: we wrote up a post on Agentic MCP Configuration that we're very excited about. A big reason Anthropic's Agent Skills concept is landing so well this week is because of its "progressive disclosure" pattern: the idea that the agent doesn't need to pull in context until it has decided it needs it. And our Agentic MCP Configuration pitch follows that same arc: you don't need to load those MCP servers until you actually learn you need them.

Cheers,
Mike and Tadas