The AI Skill Marketplace Is Coming. Here's What's at Stake.

The GPT Store failed because OpenAI owned the platform. The AI agent skills marketplace will succeed for the opposite reason. No single company owns it. The format is the platform, and the creator economy is forming around the format whether the platforms cooperate or not.

That is the bet, and it is forward-looking. Let's be honest about that up front. The monetization layer does not exist yet as of mid-2026. The infrastructure does. The infrastructure is what makes the monetization inevitable.

What Already Happened

The first attempt was the GPT Store. OpenAI announced it on January 10, 2024, with a builder revenue program attached. The promise: US builders would be paid based on user engagement with their GPTs. At launch, OpenAI reported that users had created over 3 million custom GPT versions in the two months since custom GPTs were introduced.

That was the high water mark. By approximately May 2024, the revenue program was quietly dropped. No earnings data was ever published. There was no farewell post. The store still exists. The builders mostly stopped building.

The post-mortem on the GPT Store is usually "consumers did not want it" or "the discovery was bad." Those reads are incomplete. The structural read is that OpenAI controlled the discovery surface, controlled the pricing model, controlled the payout formula, and then quietly turned the payout off. Builders were renters on someone else's platform. When the landlord changed the terms, the building emptied.

The lesson the rest of the market took from this was not "marketplaces do not work." The lesson was "platform-controlled marketplaces have a single point of failure, and that point is the platform's enthusiasm."

What Is Happening Now

The newer move is structural. Anthropic released the Skills format and made it an open standard at agentskills.io. The GitHub repository, as of June 2026, has roughly 20,000 stars and 1,200 forks. The exact launch date is not as clean as the press would suggest, but the trajectory is.

Here is the structural difference that matters. A Skill is a folder. The folder has a manifest, some instructions, and optional scripts. It is a file format. It runs in Claude Code. It runs in Claude apps. It runs through the API. The spec is published. Any tool that implements the spec can load any skill that follows the format. The Cursor team could decide tomorrow to support Skills natively, and existing Skills would just work. The VS Code extension authors could do the same.

This is the npm shape. Or the Docker image shape. Or the email shape. A format that any tool can implement, that nobody owns, that becomes valuable because the corpus of compatible content compounds across all the tools that adopt it.

The Custom GPT, by contrast, runs in one place. If OpenAI dropped Custom GPTs tomorrow, every Custom GPT would stop working at the same moment. The Skill does not have that single point of failure. The teams running on Claude Code today, with their skill libraries, could move to a hypothetical Cursor implementation tomorrow, with no migration cost beyond pointing the loader at a new directory.

That is the bet. The portability is the platform.

Who Is Earning Money Today

The honest answer: not from skills, yet. From adjacent surfaces, plenty.

Pieter Levels, the canonical indie operator, runs Photo AI and Interior AI and a few other things. Public reporting from 2023 put the combined revenue at roughly $150,000 per month. The current number is likely different and his public dashboards are worth checking before quoting any specific figure. These are agentic micro-products, not skill marketplace entries, but they are the proof point that a single creator can earn real money from AI-native artifacts when the distribution surface is right.

The Custom GPT cohort, before the revenue program ended, included thousands of creators who built complex assistants and never saw a dollar. That experiment ended with the lesson "the revenue program needs to exist for the creator economy to mature." The lesson is informative even though OpenAI walked away from it.

Hugging Face Spaces has hundreds of thousands of creator-uploaded apps and demos. The platform does not pay creators directly. Some creators monetize through Pro tiers, sponsorships, or by funneling traffic to their own products. The pattern: even without a direct payment layer, creators are publishing because the publishing itself produces career value. The monetization is a side effect.

Cursor's MDC rule files and .cursorrules ecosystem has thousands of community contributions on GitHub and in unofficial directories. Free, informal, no monetization layer. Same pattern: creators publish for career and reputation, not direct revenue. When the monetization layer arrives, the creator pool is pre-populated.

What Has To Be True for the Marketplace to Exist

Three things, in order.

Critical client mass. The Skill format needs to be supported by enough tools that a creator can publish once and reach a meaningful audience. Right now it is supported by Claude Code, Claude apps, the API, and a few third-party clients. If the number gets to ten well-known tools by end of 2026, the marketplace forms. If it stays at three, the format becomes a Claude-specific thing and the structural advantage is lost.

A monetization primitive. Someone has to ship the layer that lets a creator charge for a Skill. It can be Anthropic. It can be a third party that runs the marketplace on top of the open format. It can be Stripe with a thin shim. The first version of this layer will be ugly. The second version will get the unit economics right. The third version will look like the App Store except open.

Curation that is not gatekeeping. The GPT Store had curation that was gatekeeping. OpenAI decided which GPTs to feature. That made discovery a popularity contest run by one company. The marketplace that wins will have curation that is decentralized: ratings, reviews, social signals, embedded inside the tools where creators already publish. The discovery surface lives in the editor, the agent, the IDE, the workspace. Not in a separate destination URL that nobody visits.

If those three things land in the next 12 to 18 months, the AI agent skills marketplace exists. If only one or two land, the format becomes infrastructure for closed marketplaces (vendor lock-in repeated) rather than an open creator economy.

What This Means If You Are Building

If you are a solo founder or a small team thinking about authoring AI skills, here is the strategic read.

Publish now, monetize later. The creators who win when the monetization layer arrives are the creators with a body of work and a reputation already in place. The Hugging Face pattern. The npm pattern. The early YouTubers in 2008. Get the first ten Skills out before you need them to make money.

Pick a niche the platforms will not service themselves. Anthropic is not going to ship a Skill for invoicing in your specific vertical industry. They will not write the Skill that automates the back-office for a single-truck plumbing operation in three US states. Niche Skills are defensible because the platforms have to chase mass adoption. Solo creators can serve a category of 100,000 customers and make a good living without ever showing up on a vendor's roadmap.

Build for the format, not the platform. Write Skills that follow the open spec. Avoid Skills that depend on Claude-specific extensions or undocumented behaviors. The portability is the asset. If your Skill works in three tools, your customer base triples for free the next time a fourth tool adopts the spec.

Document your evals. The Skill that wins is not the one that markets best. It is the one that publishes its own eval rubric and shows results against it. The customers who are buying Skills are buying verifiability. Hand them the verification on a plate.

The Honest Limit

The marketplace story is forward-looking. The infrastructure is real. The monetization layer is not. Anyone telling you a specific number about Skill creator earnings in 2026 is making it up. The future shape is clear; the timing is not. If your plan depends on this market being mature by Q4, you have a planning problem, not a market problem.

The bet that is reasonable to make is on the format, not the timing. The Skill format is durable infrastructure now. The marketplace will form on top of it. Creators who position themselves on the right side of the open format are positioning for compounding leverage when the layer above ships.

The GPT Store taught us that platform-controlled marketplaces are fragile. The Skill format is teaching us that open standards do not have that fragility. The next twelve months will tell us whether the format-as-platform thesis was right.

The model is fine. The platform shape is the bet.