The Power Law of Agent Plugins: What OpenClaw Tells Us About the Future of AI Extensions

A New Kind of Platform

OpenClaw is an extension framework for AI agents, specifically, for Claude-based agents running in agentic environments like Clawdbot. Think of it as a plugin system where the host application is not a browser, a CMS, or an IDE, but an AI agent itself. Skills are plain-text instruction files that tell the agent how to behave in a given context: how to use a CLI tool, how to connect to an API, how to manage memory, how to improve itself.

The skill registry, ClawHub, is the distribution layer. As of March 2026, it hosts 31,399 skills with public download and star counts, which makes it one of the first agent plugin markets to have enough data to analyse seriously.

This has an interesting implication: the bottleneck for skill quality is not engineering but clarity. The best skills are not the most technically complex, they are the most precisely written. As we will see, this creates a very different competitive dynamic from software plugin markets that came before.

Who Dominates, and Why

To understand what comes next, it helps to look at what has already played out. The top 50 skills by download count reveal patterns that are both distinctive to this market and deeply familiar from analogous ones.

The power law in raw form

Download distributions in plugin ecosystems are not normally distributed. They are not even mildly skewed. They follow what mathematicians call a power law, a mathematical relationship where a small number of items accumulate a wildly disproportionate share of total value, and each successive rank earns dramatically less than the one above it.

In a power law distribution, the frequency of an item as a function of its rank follows: f(r) = C · r^−α, where r is the rank and α (alpha) is the exponent that determines how steep the drop-off is. The OpenClaw data shows a very steep alpha, meaning the drop from #1 to #10 is far more severe than in, say, the App Store, which has years of long-tail accumulation to moderate the curve.

The top three skills alone, self-improving-agent, ontology, and Self-Improving + Proactive Agent, account for roughly 495,000 downloads. That is more than all skills ranked 11–50 combined. Not unusual for a young market: the WordPress plugin ecosystem looked nearly identical in 2007.

The thematic map

Grouping the top 50 by use case rather than rank makes the picture clearer. The market has not yet found equilibrium across all categories, some are saturated with competing skills, while others are winner-take-all.

Stars versus downloads: the quality signal problem

Raw download counts are a noisy signal, and always have been. In every mature plugin market, npm, WordPress, Chrome Extensions, download inflation through bots, CI pipelines, and programmatic installs eventually degrades the metric's usefulness. Stars, which require a deliberate human choice, are more signal-dense.

The ratio of stars to downloads is a simple way to tell whether people who installed a skill actually cared about it. A healthy ratio sits above 0.5%. Two skills in the top 50, Polymarket (#13) and Post Job (#29), have ratios below 0.05%, suggesting their download counts may be inflated by programmatic or incentivised installs.

We Have Seen This Before

Every major software platform that opened a plugin layer went through roughly the same sequence: an initial explosion of creator activity, a rapid power-law consolidation, a quality crisis, and then a maturation phase defined by monetisation and platform-level curation. OpenClaw is early in step one. Understanding how that sequence plays out is probably the most useful thing a skill creator can know right now.

The WordPress Plugin Ecosystem (2005–2012): the closest structural analogue

WordPress plugins are the closest comparison to OpenClaw skills. Both are plain-text layers sitting on top of a larger platform. Both have almost no barrier to publishing. And both reward credentialed, prolific creators with distribution advantages that compound over time.

The numbers tell the story clearly. WordPress introduced its plugin architecture in May 2004. An official directory launched in 2006. By the end of 2011 that directory had 10,000 plugins. By early 2013, 20,000. Total downloads crossed 1 billion in 2014 and the directory passed 50,000 plugins by 2017. The growth was exponential in the early years, then began flattening as the platform matured and started pruning low-quality entries. OpenClaw at 31,399 skills is somewhere between WordPress's 2012 and 2013 levels in raw count terms, suggesting the consolidation phase is closer than it looks.

In WordPress's first three years or so, a handful of authors, Automattic insiders, Joost de Valk (Yoast SEO), Carl Hancock (Gravity Forms), built positions in the top 20 that held for five years. The meta-plugin pattern was already visible by 2007: SEO plugins, caching plugins, and security plugins dominated because they made WordPress itself more capable rather than adding a single feature. Yoast SEO is now active on over 10 million sites. It did not solve a specific problem, it solved the problem of WordPress's relationship to search engines generally.

@steipete on OpenClaw maps almost exactly to Joost de Valk on WordPress: a credentialed insider, prolific publisher, high version cadence, and deep integration with the platform's own roadmap. The moat is not any single skill, it is the trust signal that accretes across all of them.

iOS, Chrome, and npm: speed, governance, inflation

The iOS App Store launched in July 2008 with 500 apps. Within a single weekend it had 10 million downloads. By September 2012, just four years in, it had 700,000 apps and 30 billion cumulative downloads. Chrome's extension gallery grew from zero to 8,500 extensions and 70 million installs in its first year, reaching 750 million total installs by mid-2012. npm, more quietly, went from nothing in 2010 to 100,000 packages and 500 million monthly downloads by 2014, and 3.3 million packages by 2019.

Each of these markets went through the same governance arc. Early: open, fast, almost no review. Middle: malicious actors exploit the openness, platform scrambles to respond. Late: stricter policies, quality floors, pruning of low-value entries. Chrome removed extensions that injected ads without user consent. Apple tightened privacy rules and killed 64-bit non-compliant apps, trimming the store from 2.2 million to 1.9 million apps. npm responded to spam and supply-chain attacks with new policy tools. The pattern is invariant. The only question for OpenClaw is how long the open phase lasts.

Shopify: the B2B template for what comes after

Shopify launched its App Store in June 2009. Growth was slow for the first few years, then exponential after 2012. By 2021 it had 6,000 apps. By 2024 it crossed 10,000, before a modest pullback as Shopify pruned underperforming entries. The key difference from every other analogue: each install on Shopify connects to a business that generates real revenue. Apps that improve conversion or reduce operational costs can justify recurring subscription fees easily, because the ROI is direct and measurable.

This is the model OpenClaw is most likely to evolve toward if it attracts professional and enterprise users. The API Gateway skill (#8) already looks like a Shopify-style connector: it sits close to the workflow that matters, it integrates external services merchants (agents, in this case) actually pay for, and it would be a natural candidate for a paid tier. When that transition happens, the ranking dynamics could shift substantially. Shopify's experience suggests that the free-to-paid transition reshuffles the top 50 faster than any other platform event.

The power law: why concentration is mathematically inevitable

The power law is not a quirk of one market or another. It emerges from systems with preferential attachment, where new nodes in a network are more likely to connect to already-well-connected nodes. In plugin markets the mechanism is pretty simple: users sort by popularity, install the top result, leave a star if it works, which pushes it higher in the rankings, which drives more installs. The rich get richer, not through any conspiracy, but because of how the distribution is shaped.

Mathematically, if you plot rank on the x-axis and downloads on the y-axis on a log-log scale, a pure power law gives you a straight line. The OpenClaw data is not quite a straight line yet, it is steeper at the top than a mature market, reflecting the fact that the distribution has not had time to develop a proper long tail. As more skills accumulate downloads over time, the curve will flatten and broaden. But the top of the distribution will stay compressed: the #1 skill will likely always hold a multiple of 5–20× over the median.

The platform commoditisation trap

The one risk that historical analogues cannot fully capture is the one most specific to AI: the underlying platform is itself improving at a rate that dwarfs any software platform in history. WordPress core improved incrementally, a new block editor here, better REST API there. The LLMs powering OpenClaw agents are on a trajectory where capabilities roughly double every 12–18 months.

This creates a kind of risk that the App Store, npm, and WordPress never had to deal with: the skills that currently occupy the top 3 positions, all agent self-improvement skills, are precisely the category the platform itself has the most incentive to absorb. If Claude 4 or 5 natively handles persistent memory, self-correction, and context management, the top 3 skills on ClawHub get deprecated by a model update. This is the equivalent of Apple folding flashlight apps into iOS, except it would happen to the highest-value category on the platform rather than a minor utility.

The Historically Informed Playbook

Everything that follows is derived from pattern-matching against the historical analogues above, filtered through the specific dynamics of the OpenClaw data. These are not generic startup tips. Each one maps to something specific in the data.

Understand control points before you pick a category

One thing the Bain and McKinsey research on platform ecosystems makes clear is that the biggest long-term winners in plugin markets are not the ones who picked the most popular category. They are the ones who positioned themselves near a control point: a part of the platform stack that is hard to replicate and that many other participants depend on.

In WordPress, the control points were the official directory, the update mechanism, and the default bundled plugins (Akismet shipped with every install). In iOS, Apple owned distribution, billing, and device APIs as three interlocking control points. In npm, the central registry URL itself became a control point: everything routed through it, making certain packages critical infrastructure. In Shopify, checkout and the merchant billing relationship were the control points that made every app a complement rather than a competitor.

On OpenClaw the control points are still forming, which is exactly why now is the right time to think about them. The most defensible positions will be: the skill that owns the discovery layer (a meta-skill that manages and recommends other skills), the connector layer (the skill through which agent workflows touch external services), and the trust and verification layer (the skill that tells you which other skills are safe to run). Two of those three slots already have credible occupants. The third, discovery, is wide open.

Conclusion: this is 2007 for agent plugins

Every signal in the OpenClaw data is consistent with a market in its first 18 months. The power law is steep and still forming. Creator moats are crystallising in real time. The quality signal problem is visible already, just not critical yet. Monetisation has not arrived yet. The platform is improving faster than any analogous historical market ever did.

That combination of familiar dynamics and novel risk means the opportunity is real and the window is open, but it will not stay open long. The top-10 list in most analogous markets became nearly immutable within two to three years of launch. On OpenClaw, given the pace of the ecosystem, that window may be shorter.

The people who win will be the ones who treat this not as a novelty but as a familiar historical pattern playing out in a new context, and act with the urgency that kind of pattern recognition demands.