In 2026, the distribution of extensions for AI agents has finally taken shape around the "pack": the universal capabilities of models have been devalued, and the sets of skills, commands and tools gathered together have become a product. Major players have structurally fixed this shift: xAI launched the Grok Build plugin marketplace with a catalog in the form of an open repository on GitHub, and Anthropic transferred the MCP protocol to the neutral Agentic AI Foundation under the Linux Foundation umbrella. The author of the Telegram channel assesses both steps as confirmation of his thesis "generics are a commodity, pack is a product", although the scale of this conclusion raises reasonable doubts.

image
image

What happened

On June 11, 2026, xAI launched a plugin marketplace for the Grok Build agent. A plugin here is a single package combining skills, slash commands, agents, hooks, as well as MCP and LSP servers. The catalog became a public storefront on GitHub, the xai-org/plugin-marketplace repository: authors submit plugins via a regular pull request, and remote plugins are pinned to a specific commit SHA, which fixes the version of the tool. Users received two ways to install: a storefront directly inside the Grok Build terminal interface and an installation command from the CLI, with installation by default proceeding with an explicit trust flag. At the start, the catalog features MongoDB, Vercel, Sentry, Chrome DevTools, Cloudflare and Superpowers. The infrastructure background is the December 2025 step: Anthropic, together with Block and OpenAI, transferred the MCP protocol to the Agentic AI Foundation under the auspices of the Linux Foundation — according to the analysis, goose and AGENTS.md also went there.

Context

MCP is a protocol for connecting external tools to models, originally proposed by Anthropic; an ecosystem of tool servers has grown around it, and together with the SKILL.md specification and Claude Code formats, a de facto standard for packaging agent capabilities has emerged. Storefronts for such extensions until recently were built as closed curated catalogs: a telling example is Meta's Connector Directory, which operates without public documentation on the rules of entry. xAI's approach is the opposite: the catalog is declared a regular git repository, where you get in via pull request. The neutral jurisdiction of MCP at the same time solves the classic problem of platform standards: a protocol belonging to one vendor always raises the question of capture among competitors, while a foundation under the Linux Foundation removes it. Against this backdrop, the channel author's thesis sounds like this: "generics" — the basic capabilities of models — have become a commodity, available everywhere, and the product for sale turned out to be precisely the pack.

Why this matters for the industry

For the industry, the very unit of distribution is being consolidated: an extension is already described not as a library, but as an aggregate of contracts — skills, slash commands, agents, hooks, MCP and LSP servers in one package. The cost of market entry for tool authors drops to almost zero: instead of negotiations with the storefront — a regular pull request, and the pinning of remote plugins to commit SHA provides a reproducible version of the environment for experiments and production environments. Platform competition shifts from interfaces to entry rules: xAI's open PR catalog against a closed curated funnel like Meta's Connector Directory; Muse has no storefront at all, and closed schemes without public rules remain a weak point of this scenario. There is also a side effect of the aggregate nature of the plugin: complexity shifts from the model to the environment, so in agent benchmarks without additional control it is difficult to separate the capabilities of the model itself from the capabilities of the installed pack. Monetization of "pack as a product" through query-time keys and metering — this is the channel author's construction, not a confirmed market practice.

Why this matters for users

For those who already publish skills or plugins in the Claude Code or SKILL.md format, according to the official documentation, there is no need to redo anything: Grok Build reads these same formats (skills, plugins, marketplaces, MCP, hooks), and installation for the user is reduced to the "/marketplace" storefront inside Grok Build or the "grok plugin install --trust" command in the terminal. The initial set of the catalog is made up of applied connectors to well-known development services, so the usefulness of the storefront is immediately visible, without choosing from an empty list. For authors of research artifacts — evaluators and connectors — this is a simple path to an audience without their own site and storefront, and for teams — a mechanism for delivering toolkits to working environments with a fixed version.

What is still unknown / limitations

The main caveat concerns the scale of the thesis: "pack is a product" is presented as confirmed by the market, but it is based on observations of several major players, and several companies with pack distribution fix a trend, not prove causality; the channel author's own architecture coincides with this thesis, so the source has a promotional bias. There is no scientific novelty in the events: this is infrastructure convergence, not a new architecture, training method or benchmark. Grok Build's compatibility with Claude formats "without rework" is based on the documentation and reviews cited in the post; there is no independent verification on non-trivial packs yet, so this is a working hypothesis. Installation proceeds with the --trust flag in the absence of a gatekeeper, and the rules of trust in third-party plugins remain an open question. There is also no independent data on the adoption of the catalog — the volume and quality of plugins, real cases of incompatibility — yet.

Sources

Author

Look at AI, editorial team