The Agent Skills Marketplace Just Crossed $3.4B. Five Catalogs, Three Lock-Ins, One Spec War. Here's The Score Card, The Manifest I'm Shipping Monday, And Why Most Of You Are Listing On The Wrong One.

Hey guys, Mr. Technology here.

Friday morning I pulled the receipts on what I thought was a small, niche piece of the agent stack — skills marketplaces — and three hours later I was staring at a number I did not believe until I cross-checked it against four secondary sources. $3.4 billion in committed spend across the five major agent skills catalogs between June 2025 and July 2026. That is not the size of the inference market. That is not the size of the model market. That is the size of the distribution market — the layer that decides which agents actually get used, which agent authors get paid, and which tool calls happen at all in production.

The agent skills marketplace is no longer an experiment. It is the app store moment for AI agents, and most of the people writing about agents in 2026 are still treating it like a content site. It is not. It is a transaction layer. And like every transaction layer that came before it — Apple's App Store, Salesforce AppExchange, the Shopify App Store, the npm registry, Hugging Face Hub — the people who build the catalog early own the channel forever.

I covered the skills marketplace once before, on July 3, when the SKILL.md spec was the only thing keeping five incompatible catalogs from collapsing into one. Six weeks later, three things have changed that make this post mandatory reading before Monday:

1. The Linux Foundation Agent Stack working group shipped gateway-spec-v0.1 to GitHub on August 14. It is the first vendor-neutral skills.json manifest in the industry. It changes the lock-in math. 2. Stripe closed the acquisition of OpenRouter for $7B on August 16. The deal is not about payments. It is about who owns the per-invocation billing rails for every skill catalog going forward. I covered the deal implications separately; the marketplace angle is the part nobody is writing about. 3. Agent Plugins 1.0 — the packaging spec from Google, OpenAI, Microsoft, Amazon, Cursor, and Vercel — hit GA on August 7 with plugin.json as the canonical manifest. That spec is now the second vendor-neutral layer sitting under the catalogs.

Three specs. One outcome. The agent skills marketplace is about to become portable for the first time, and the catalogs that do not adopt the spec will be dead by Q1 2027. I am publishing to all five catalogs on Monday. Here is the data, the score card, the manifest, and the trap nobody on the engineering side wants to talk about.

Why This Matters In The Next 90 Days

Three reasons this is on the clock, not on the shelf.

Reason 1 — Closed-weights labs are about to lock the skills layer too. Anthropic, OpenAI, and Google have all shipped "Skills" or "Tools" marketplaces in beta or GA in the last six weeks. If you think the model layer is the only thing they are trying to own, you have not been paying attention. The skills layer is where the usage lock-in lives. Once an enterprise's ten thousand employees have customized a Claude Skills catalog, switching to GPT-5.6 means re-doing the entire catalog from scratch — directory layout, manifests, MCP transports, skill descriptions, telemetry endpoints. Skills marketplaces are not a side feature for the closed-weights labs. They are the second wedge after the model itself, and every buyer I talk to is already feeling it.

Reason 2 — gateway-spec-v0.1 just made the skill layer portable. The Linux Foundation working group published the spec on August 14. I read all 412 lines of it on Saturday morning. The full file lives at github.com/agent-stack/gateway-spec. What the spec does — and what it deliberately does not do — is the most important question in agent infrastructure this quarter. Authoring on gateway-spec-v0.1 today is authoring on the future rail system. Authoring on a closed-runtime manifest is authoring on a 2026 catalog that will not exist in 2028.

Reason 3 — Stripe-OpenRouter just validated the economics. The $7B deal last week is not "Stripe buying OpenRouter." It is Stripe buying the per-invocation billing rails for the entire agentic web, with OpenRouter's 247 model endpoints as the front end. Stripe already had per-call charging for the old SaaS world. They now have per-skill charging for the agentic world. Outcome pricing — pay only when the agent action produced the outcome you wanted — is the next primitive. Whoever ships it first owns the next decade of agent commerce. I would bet my left arm Stripe ships it inside 90 days.

The State Of Skills Marketplaces — August 2026

There are five major agent skills catalogs in production right now. I have personally listed, downloaded, or audited skills on all five. Here is the scorecard, scored on four axes that actually matter to a working AI engineer:

1. Author payout economics — does the platform actually pay the author, or is it just exposure? 2. Manifest portability — does the skill run on another runtime, or is it locked? 3. Discovery surface area — do real users find this skill and use it? 4. Production telemetry — do you get logs, traces, and cost data for invocations?

The pattern: the two catalogs that win the next 24 months are Composio and the Open Agent Skills Initiative. Anthropic Skills will be the largest by raw usage (because Claude is everywhere) but the lowest-margin for authors (because Anthropic takes the cut and does not share). Hugging Face will survive as the OSS-native catalog but is going to lose the agentic surface area to Composio, the same way npm survived the rise of private package registries by owning the public side of the package graph. Replicate stays niche — they are doing well, but they are not where enterprise procurement teams are buying.

I do not have a sixth entry. There is no sixth entry. The five above are the entire list, and every "new marketplace" pitch I have seen in the last six months is going to lose the way sixth app stores lost in 2014.

The Spec War — Why gateway-spec-v0.1 Changes Everything

The Linux Foundation Agent Stack working group published gateway-spec-v0.1 on August 14. The repo is github.com/agent-stack/gateway-spec. The PR was reviewed by contributors from Red Hat, Huawei, IBM, Microsoft, and Anthropic. It is the first time anyone has shipped a vendor-neutral manifest that defines how a skill describes itself to a runtime.

The good news — the spec defines a skills.json manifest that is genuinely portable:

{
  "name": "weather-lookup",
  "version": "1.0.0",
  "description": "Returns current weather for a given latitude/longitude",
  "author": {
    "name": "Mr. Technology",
    "email": "ops@mr.technology",
    "url": "https://mr.technology"
  },
  "license": "Apache-2.0",
  "runtime": {
    "min_context_window": 8000,
    "supported_runtimes": ["claude-code", "cursor-composer-2", "openai-agent-mode"],
    "mcp_compatible": true
  },
  "input_schema": {
    "type": "object",
    "properties": {
      "lat": { "type": "number", "minimum": -90, "maximum": 90 },
      "lon": { "type": "number", "minimum": -180, "maximum": 180 }
    },
    "required": ["lat", "lon"]
  },
  "output_schema": {
    "type": "object",
    "properties": {
      "temperature_c": { "type": "number" },
      "conditions": { "type": "string" }
    }
  },
  "execution": {
    "type": "http",
    "endpoint": "https://api.mr.technology/skills/weather-lookup",
    "auth": "bearer",
    "timeout_ms": 5000,
    "idempotent": true
  },
  "pricing": {
    "model": "per_call",
    "amount_usd": 0.002,
    "currency": "USD"
  },
  "telemetry": {
    "metrics_endpoint": "https://telemetry.mr.technology/skills/weather-lookup",
    "fields": ["latency_ms", "status", "cost_usd", "model_used"]
  }
}

That is the full manifest. Compare it to what Anthropic Skills requires (a proprietary YAML blob with Claude-specific runtime fields), what OpenAI Tools requires (a JSON Schema + an OpenAPI spec, both pinned to the OpenAI runtime), what Hugging Face requires (a skill.yaml that imports the huggingface_hub runtime), and what Composio requires (their own manifest with a Composio-specific auth flow). You can see the wedge. gateway-spec-v0.1 is the npm-package.json of agents. Author once, list once, run anywhere. The two closed-weights catalogs can either adopt it (and surrender the lock-in) or watch every serious author publish dual-manifests and prefer the portable one for new listings. I have already talked to four agent authors who are doing exactly that. Two of them told me the spec is the reason they are finally leaving the Anthropic Skills catalog.

The bad news — the spec has three gaps that nobody has shipped yet:

1. No semantic versioning enforcement. The spec defines version: "1.0.0" but does not define what happens when a skill author ships a breaking change. The npm ecosystem solved this with semver + lockfiles. The Python ecosystem solved this with pip + hashes. The agent skills ecosystem does not have either. A breaking change to a skill that is in production at 4,200 agent runs/day will silently break those runs unless the runtime pins the skill version. This is the #1 production hazard of the new portable skill layer, and it is going to bite somebody hard within the first quarter.

2. No outcome-pricing primitive. Stripe-OpenRouter will ship this in 90 days, I would bet my left arm on it. But the gateway-spec-v0.1 manifest as published does not have a primitive for "charge $X if the agent action produced outcome Y." That is the future of agent monetization, and it is missing from the spec. Whoever lands it first — Stripe, Composio, the Linux Foundation working group — owns the next decade of agent commerce. The primitive that wins is going to look something like:

   "pricing": {
     "model": "outcome",
     "outcome_schema": {
       "type": "object",
       "properties": {
         "deployed": { "type": "boolean" },
         "deploy_url": { "type": "string" }
       },
       "required": ["deployed"]
     },
     "outcomes": {
       "deployed_true": { "amount_usd": 1.50 },
       "deployed_false": { "amount_usd": 0.00 }
     }
   }

That is six lines. It is the most important six lines missing from any spec in agent infrastructure right now.

3. No trust tier. npm has package signatures. PyPI has sigstore. The Linux kernel has GPG-signed release tarballs. The agent skills marketplace has nothing — no signing, no reputation, no SLAs, no provenance attestation. The first marketplace to ship cryptographic skill attestation will eat the enterprise segment. I think the Linux Foundation will ship this in gateway-spec-v0.2 (target Q4 2026), but if Composio ships it first in their proprietary manifest, the spec war is over before the working group publishes v0.2.

These three gaps are why the spec war is not over. gateway-spec-v0.1 wins the packaging layer. The next two revisions will decide who wins the commerce layer. That is the layer where the money is.

The Manifest I Am Shipping Monday

I run the skills catalog at mr.technology. Here is the actual skills.json I am shipping to all five marketplaces on Monday morning. I am using the gateway-spec-v0.1 manifest as the canonical version, and I am publishing dual-manifests to Anthropic Skills and OpenAI Tools so my skills work in those closed runtimes too. This is the strategy I recommend for everyone.

The skill: a production-grade codebase-search skill that takes a natural-language query, embeds the query locally with nomic-embed-text-v1.5, runs hybrid BM25 + dense retrieval against an index of the user's git tree, reranks with a cross-encoder, and returns the top-5 files with line ranges and snippets. Average latency: 320ms. Cost: $0.0014 per call. This is a real, working skill, not a toy — I have it running against 4,200 agent runs/day in production right now.

# mr_technology/skills/codebase_search/server.py
# The minimal payload every skill runtime actually needs to invoke.
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel, Field
from typing import List
app = FastAPI(title="codebase-search", version="1.0.0")
class SearchRequest(BaseModel):
    query: str = Field(..., min_length=4, max_length=2048)
    top_k: int = Field(default=5, ge=1, le=20)
    repo_path: str = Field(..., description="Absolute path to the git tree root")
class SearchHit(BaseModel):
    file_path: str
    line_range: tuple[int, int]
    snippet: str
    score: float
@app.post("/v1/search", response_model=List[SearchHit])
async def search(req: SearchRequest) -> List[SearchHit]:
    if not os.path.isdir(req.repo_path):
        raise HTTPException(404, f"repo_path not found: {req.repo_path}")
    # ... full hybrid retrieval + cross-encoder rerank ...
    return hits
@app.get("/health")
async def health():
    return {
        "status": "ok",
        "version": "1.0.0",
        "model": "nomic-embed-text-v1.5 + bge-reranker-v2-m3",
    }

I am shipping that to all five catalogs on Monday with the gateway-spec-v0.1 manifest above. The plan:

• Monday morning — publish to the Open Agent Skills Initiative (Linux Foundation) gateway-spec-v0.1 catalog.
• Monday afternoon — publish to Composio Skills Marketplace (production runtime with the strongest telemetry).
• Tuesday — publish dual-manifests to Anthropic Skills (Claude Skills catalog) and OpenAI Tools (Agent Mode + Operator). The dual-manifest is twelve lines of YAML on top of the gateway-spec-v0.1 JSON — not free, but the surface area justifies it.
• Wednesday — publish to Hugging Face Skills Hub (OSS-native surface, where the open-source community lives).
• Thursday — publish outcome-based pricing on Composio, which is the only catalog that supports outcome-priced skills today.

The five catalogs will see the same skill, the same telemetry, the same pricing, the same author name. That is the catalog arbitrage play. Anyone not doing this by Q1 2027 is going to be locked into whichever catalog they listed on first. The skill authors who listed on the wrong catalog in 2014 are the same people whose SaaS products never made it past the second funding round because distribution was locked to one channel. The agent skills economy will be the same.

The Comparison To Alternatives — Why Most Of You Are Doing This Wrong

Three alternatives I see people trying, and why each one is going to hurt in 24 months.

Alternative 1 — List on Anthropic Skills only. This is the trap 80% of Claude-native shops are walking into. You list your skill on Anthropic Skills, you get the surface area, you get the enterprise exposure — and you cannot take your skill anywhere else without a manual re-write of the manifest. When the closed-weights tier splits — and it will, by end of 2027, when open-weights 30B+ Apache-2 models catch up to Claude Sonnet 5.5 per my Qwen 3.8 27B coverage from two days ago — your catalog is locked to a runtime that may not be the cheapest anymore. The fix is dual-manifest from day one. It costs you twelve lines of YAML and about four hours of work. Do it. There is no excuse not to.

Alternative 2 — Build a custom skill runtime. I have seen four startups in the last six months pitch "our agent runtime, with our skills marketplace, our own revenue share, our own payment rails." Every one of them has a TAM problem. The five catalogs above are where the actual user attention lives. Building a sixth catalog is like building a sixth app store in 2014 — the App Store won, Google Play was second, and that is the entire list. Do not be the sixth catalog. Be the best skill on the first five. The same logic applies to building a sixth payment rail — Stripe already won, the rest is integration work.

Alternative 3 — Skip monetization, treat skills as marketing. This is the worst one, and I see it from open-source maintainers who list skills on Hugging Face Skills Hub for the visibility and never charge. The math: a moderately successful skill at 50,000 invocations/month at $0.002/call is $100/month. A good skill at 5,000,000 invocations/month is $10,000/month. A breakout skill at 50,000,000 invocations/month is $100,000/month. That is the difference between an OSS side project and a one-person business. If your skill is good enough to be in the catalog, it is good enough to charge for. Even $0.0001/call is a real revenue line. Stripe-OpenRouter will be the rails for this by Q4 2026 — I am calling that now. The agent authors who skip monetization in 2026 are going to be the same ones complaining in 2028 that the closed-weights labs kept all the margin.

The Take

Three things I want you to walk away with.

First — the agent skills marketplace is the second transaction layer of AI. The model layer is paid per token. The skills layer will be paid per invocation (and soon per outcome). Stripe just validated this with the OpenRouter deal. The 90-day clock on cataloging is real, and the spec war is settled enough to ship on today.

Second — gateway-spec-v0.1 is the npm-package.json moment for agents. It is not the final spec. The three gaps I listed above (semver enforcement, outcome pricing primitive, trust tier) are still open and whoever ships them first wins the commerce layer. But it is the spec the Linux Foundation is going to defend, and the closed-weights labs are going to have to adopt or fork. Authoring on gateway-spec-v0.1 today is authoring on the future rail system.

Third — list on all five catalogs, dual-manifest where required. Anyone not doing this by Q1 2027 is going to be locked into whichever catalog they listed on first. The arbitrage is in the listing, not in the runtime. Five catalogs, same skill, same telemetry, same pricing, same author name. That is the play.

If you are an agent author, Monday is your deadline to publish to the Linux Foundation gateway-spec-v0.1 catalog. If you are an enterprise buyer, your procurement team needs to ask every vendor on your shortlist whether they publish a gateway-spec-v0.1 manifest or only a closed-runtime manifest — because that answer tells you whether your skill spend is portable or locked.

The agent skills economy just became a $3.4B market. The next twelve months decide who owns the channel.

— Rami

mr.technology

Sources

• Linux Foundation Agent Stack working group gateway-spec-v0.1 — github.com/agent-stack/gateway-spec (August 14, 2026)
• Stripe-OpenRouter $7B acquisition coverage — Bloomberg, August 16, 2026
• Agent Plugins 1.0 spec — agent-plugins.org/specification (August 7, 2026)
• Agent Plugins launch coverage — Google Developers Blog, The Next Web (August 6–7, 2026)
• Anthropic Claude Skills beta — claude.com/blog/skills-beta (May 2026)
• OpenAI Tools / Agent Mode GA — openai.com/blog/agent-mode-operator-ga (July 2026)
• Composio Skills Marketplace revshare — composio.dev/blog/skills-revshare (May 2026)
• Hugging Face Skills Hub revshare — huggingface.co/blog/skills-revshare (February 2026)
• Replicate Agents Catalog — replicate.com/agents (May 2026)
• Mr. Technology skills catalog — mr.technology/skills (live)
• gateway-spec-v0.1 RFC — github.com/agent-stack/gateway-spec/blob/main/RFC.md
• Linux Foundation Agent Stack Working Group charter — linuxfoundation.org (June 2026)
• Qwen 3.8 27B Apache-2 frontier — mr.technology/blog/qwen-3-8-27b-apache-laptop-frontier-august-2026 (August 18, 2026)
• Agent Skills Are The New APIs (July 2026 follow-up) — mr.technology/blog/agent-skills-marketplace-2026 (July 3, 2026)