A marketplace for agent skills — codified workflows, evaluation harnesses, taste constraints, and tool wiring — that people install into Cursor, Claude Code, or any agent host. The buyer gets someone else’s way of getting good outcomes from an agent. The seller earns revenue from codified expertise.
V1 proposed the wedge as “AEO/GEO design taste.” This R2 stress-tests that framing against live market data.
Eric San already runs this in production. His PCRM workspace contains 24+ skills across research, social, writing, and operational workflows.1 He spends ~US$3,400/mo on Cursor2 and has built the most integrated personal agent system we’ve seen. His .cursor/skills/ folder is the prototype product. The question is whether anyone else will pay for it.
SkillsMP aggregates 200K+ skills from GitHub repositories and hit 958,542 monthly visits in January 2026 — an 18x increase from 53,169 in November 2025.3 cursor.directory has 72.6K+ community members sharing rules and MCP servers, all free.6 cursor.store curates MCP servers with featured placement (paid), but the core listings are free.7
Cursor’s official marketplace bundles MCP servers, skills, rules, subagents, and hooks into installable plugins.8 Launch partners include Stripe, AWS, Figma, Linear, and Vercel. Plugins are Git repos, manually reviewed. No pricing model for third-party plugins has been announced — all current plugins appear free.9
The upstream MCP Registry provides metadata but explicitly delegates curation to downstream aggregators.10 This creates a theoretical niche for taste-layer curators. But the actual platforms filling that niche are all free (SkillsMP, cursor.directory, cursor.store).
Apify pays $500K+/month to MCP server developers (80% creator share, pay-per-event model) across 7,000+ tools and 36,000 monthly active developers.4 But Apify sells functional tools that scrape, transform, and connect — not “taste” or “workflow guidance.” This distinction matters.
| Layer | Size | Source | Honest caveat |
|---|---|---|---|
| Global: Agentic AI tools | US$10.4B (2025), 56% CAGR to $61.5B by 2029 | Research and Markets11 | This is the broadest possible category. Most of this is enterprise platforms, not skills marketplaces. |
| Segment: Template/plugin marketplaces | US$1.3B total author earnings (Envato lifetime); Shopify apps avg $93K/yr | Shutterstock/Envato12, Market Clarity13 | Established markets with 10+ years of compounding. New entrants face extreme concentration (0.24% of ThemeForest themes crossed $1M). |
| Segment: Agent skills discovery | 1.6M combined monthly visits (Jan 2026), from near-zero in Oct 2025 | AgentPuter analysis3 | Traffic ≠ revenue. Every platform in this category is free. No proven willingness to pay. |
| Regional: Cursor ecosystem | ~US$1B ARR, $29.3B valuation | Agent Tooling report2 | Cursor’s marketplace is brand new. No third-party monetization exists yet. |
| Local: AEO/GEO practitioner budget | 12% of digital marketing budgets (enterprise avg) | Conductor survey14 (vendor — see caveat below) | Vendor survey of self-selecting enterprise CMOs already invested in AEO/GEO. AI referral traffic is only ~1% of web visits today.15 |
| Addressable: Eric-specific | 68 Dunbar contacts, research.ericsan.io audience, Cursor power-user network | PCRM state1 | Tiny. No organic distribution channel for a marketplace. Eric’s network is strong for consulting, not for marketplace liquidity. |
| Player | Model | Scale | Revenue | Your opening |
|---|---|---|---|---|
| SkillsMP3 | Free aggregator from GitHub repos | 200K+ skills, 958K visits/mo | None visible | Dominates discovery. No curation or QA. You can’t out-aggregate them. |
| cursor.directory6 | Free community sharing | 72.6K+ members, 80 GitHub contributors | None visible | Community-driven. Free. Established brand in Cursor ecosystem. |
| Cursor Marketplace8 | Official plugin marketplace | Curated partners (Stripe, AWS, Figma, etc.) | No third-party monetization yet | Platform owner. If they add paid plugins, they own distribution. If they don’t, there’s no revenue channel. |
| Apify MCP4 | Pay-per-event, 80% creator share | 7K+ tools, 36K monthly active devs | $500K+/mo payouts | The proof that agent tool creators CAN earn. But they sell functional scrapers/connectors, not taste. |
| localskills.sh17 | Team skill sharing via CLI | Public beta | Not disclosed | B2B angle (SSO, SAML, analytics). Early. Could become the “npm for skills.” |
| PromptBase5 | Prompt marketplace | 260K prompts, 425K users (self-reported) | Undisclosed. Select tier: $1/download for creators | Proves the price floor: $1–7 per prompt. PromptBase Select offers $1/download — brutal economics. |
| GPT Store18 | Revenue sharing (limited pilot) | 3M GPTs created, 159K active | $100–500/mo typical; US-only pilot | Monetization promise mostly unfulfilled. Real money is B2B consulting ($5K–20K), not store revenue. |
| Company | Revenue/Scale | Playbook | Why it doesn’t apply |
|---|---|---|---|
| Shopify App Store13 | 14K+ apps, avg $93K/yr | Embedded in the platform merchants already pay for. Apps add revenue to merchants. | Shopify apps generate measurable GMV. Skills don’t have a measurable revenue attribution to the buyer. Also: Shopify median is $725/mo, 80% merchant churn. |
| ThemeForest/Envato12 | $1.3B total author payouts (lifetime); acquired for $245M | 10+ years of marketplace compounding, strong SEO moat, trust signals (reviews, sales count). | 76% of themes earn <$1K/mo. Avg theme = $17K/yr. 0.24% crossed $1M lifetime. Extreme concentration. New entrants need 400+ hours + $5K marketing.19 |
| Apify4 | $500K+/mo creator payouts, 36K active devs | Infrastructure-as-a-service: handles hosting, billing, scaling. Creators build functional tools (scrapers, connectors). | Apify tools DO things. Skills are text files that GUIDE agents. Apify can charge per-event because events are measurable. “Taste” has no event metric. |
| Notion Templates20 | Top creators $80K+ total, some $2K/mo | Simple, specific templates sold via Gumroad/Notion Marketplace. SEO + social distribution. | Notion templates are visual — buyers see the outcome before buying. Skill files are invisible until installed and run. Trust deficit is structural. |
| Company/Category | What happened | Lesson |
|---|---|---|
| GPT Store monetization18 | 3M GPTs created, but only 159K active. Revenue sharing is US-only limited pilot. Typical: $100–500/mo. Creators pivoted to B2B consulting. | Even OpenAI with 200M+ users couldn’t make “sell packaged AI workflows” work at scale. The store is a discovery layer, not a revenue engine for creators. |
| PromptBase Select5 | Launched $14–19/mo subscription, creators earn $1/download. Zero-fee selling option added. Daily USDC payouts — chasing crypto audience. | The price floor for “text-based workflow guidance” is $1–7. At $1/download, a creator needs 10K downloads/mo to earn $10K. PromptBase has never disclosed revenue. |
| “Wrapper Wars” collapse (2023–2025)21 | Wave of prompt/AI wrapper startups raised on hype, shipped “the OpenAI playground with dark mode,” died when models improved. Builder.ai: $1.5B valuation, $445M raised, bankrupt. | If the underlying platform ships your feature, you die. If AI models improve enough that default outputs are “good enough,” the taste gap evaporates. |
| VS Code extensions22 | 40K+ extensions, zero paid monetization path. Microsoft never built payment support despite developer requests. | IDE platforms have a strong cultural expectation that extensions are free. Cursor inherits this norm from VS Code. Skills/plugins may face the same ceiling. |
| Metric | Benchmark (winner) | Benchmark (average) | Our estimate | Source |
|---|---|---|---|---|
| Skill price | $50–100 (ThemeForest themes) | $1–7 (PromptBase) | $5–25 per skill | 195 |
| Take rate | 5% (Webflow) | 15–25% (Shopify, Etsy) | 15–20% (if marketplace) | 2313 |
| Subscription ARPU (alt.) | $199–500/mo (Loman AI agents)24 | $14–19/mo (PromptBase Select) | $19–49/mo (pack subscription) | 5 |
| Creator earnings ceiling | $80K total (top Notion creator) | $725/mo median (Shopify apps) | $500–2K/mo realistic | 2013 |
| Cost component | Per-unit cost | Assumption | Source |
|---|---|---|---|
| Skill creation (Eric’s time) | 4–8 hours / ~US$200–400 | At Eric’s implied hourly ($50/hr from Sourcy contract math) | PCRM state |
| Hosting / distribution | ~$0 | Skills are text files on GitHub. Cursor marketplace is free to publish. | 8 |
| QA / testing per skill | 2–4 hours / ~$100–200 | Run skill across 3–5 test cases, verify outputs, document failure modes | Estimate from PCRM skill iteration pattern |
| AI inference for testing | $1–5 per skill test round | 5 test runs at $0.20–1.00 per run (Claude/GPT-4.1) | API pricing |
| Ongoing maintenance | 1–2 hours/quarter | Model updates break skills; prompts need version bumps | Observed in PCRM skill maintenance |
| Signal | Finding | Implication |
|---|---|---|
| DEMAND | SkillsMP: 18x growth in 3 months (53K → 958K visits).3 APAC-dominant (35% China, 10% TW, 5% SG, 4% HK). | Explosive demand for skill discovery. But discovery ≠ purchase. |
| REVENUE | Apify: $500K+/mo payouts across 7K tools.4 80% creator share. Pay-per-event. | Functional MCP tools DO earn. But Apify sells scrapers and connectors — tools that DO things, not instructions that GUIDE things. |
| PRICE FLOOR | PromptBase Select: $1/download for creators. $14–19/mo subscription for buyers.5 | Price ceiling for static prompt files is brutal. PromptBase still hasn’t disclosed revenue after 3+ years. |
| TRAFFIC | AI referral traffic: ~1% of total web visits, growing ~1%/month. ChatGPT = 88.5% of AI referrals.15 | AEO is real but tiny. “Agents discover your skill page” is years away from meaningful volume. |
| ENTERPRISE | Gartner: only 23% of enterprises have optimized for AI discoverability (55-point gap).16 | Gap is real but it means buyers aren’t ready yet, not that they’re about to rush in. |
| FREE NORM | VS Code: 40K+ extensions, zero paid monetization path. Microsoft never added payment support.22 | IDE extension ecosystems establish “free is normal.” Cursor inherits this from VS Code culture. |
Skills: Build production-grade Cursor skills (proven: 24+ in PCRM). Write research reports. Operate multi-agent systems. Build HTML artifacts. Deep Cursor/Claude Code expertise.
Network: 68 Dunbar contacts (Core: 51, Warm Pro: 17). research.ericsan.io as portfolio. Strong HK tech/startup network. Sourcy + Talent Coop as active client relationships.
Bandwidth: 3–4 active projects. Sourcy at 10–20% energy cap. Donna at #1 priority. No spare capacity for a new marketplace project.
Eric’s 24+ production skills are genuinely differentiated — they integrate research, social, writing, operational workflows with tool wiring and evaluation harnesses. Nobody else has published skills at this depth. But this advantage is in the skills themselves, not in a marketplace wrapper.
| Programme | Amount | Relevance |
|---|---|---|
| Cyberport AI Subsidy Scheme (AISS)25 | Up to 70–90% of AI compute costs | Relevant for compute-heavy skill testing/inference. HK$3B fund. Applications open year-round. |
| HKSTP Co-Acceleration26 | Up to HK$15.6M equity investment | Requires HKSTP tenancy. Overkill for a skills marketplace at this stage. |
| Public Sector Trial Scheme27 | Up to HK$1M | For Cyberport incubatees. Would require Cyberport programme entry first. |
“Build a taste-layer curator for agent skills. Start with AEO/GEO design/copy workflows. Distribution: AEO-native pages that agents can cite.”
| # | Assumption | Evidence | Verdict |
|---|---|---|---|
| 1 | AEO/GEO is a real buyer with budget | Conductor (vendor): 12% budget. Gartner (independent): only 23% have optimized for AI discoverability.16 AI referral = 1% of traffic.15 | WEAKENED |
| 2 | “Taste” is defensible | AI models improve every release. GPT-5.2, Claude 4.6 — default outputs are increasingly good. The gap between “default” and “skilled” shrinks with each generation. | THREAT |
| 3 | Cursor Marketplace stays open/neutral | Cursor just launched; adding curation, ratings, paid plugins is the obvious next step. Private team marketplaces are “coming soon.”8 | THREAT |
| 4 | Users will pay for text-file skills | PromptBase: $1/download.5 GPT Store: $100–500/mo typical.18 VS Code: zero paid extensions.22 200K+ free skills on SkillsMP.3 | THREAT |
| 5 | MCP Registry stays unopinionated | Apify already provides curation + monetization on top of MCP.4 cursor.store provides featured placement. The “curation gap” is closing fast. | WEAKENED |
| 6 | AEO distribution works for skills | AI referral = 1% of web visits.15 Only 33% of consumers think GenAI rivals search.28 AEO as a distribution channel is years away from meaningful volume. | THREAT |
| 7 | Eric has bandwidth | 3–4 active projects. Donna #1, Wenhao #2, Talent Coop #3, Sourcy 10–20%. Token crisis ($200–300/day).1 | THREAT |
| 8 | Price ceiling supports a business | PromptBase: $1–7. Notion templates: $5–49. ThemeForest themes: $50 avg. Functional MCP tools on Apify: per-event billing, higher ceiling. | WEAKENED (for static skills) / HOLDS (for functional tools) |
| 9 | Enterprise practitioners buy individual skills | Enterprise buys platforms (Conductor, Semrush), not individual prompt files. localskills.sh targets team-level subscription, not skill-level purchase. | THREAT |
| 10 | “Skills” ≠ “prompts” in buyer perception | SkillsMP uses “skills.” Cursor uses “plugins” and “skills.” PromptBase uses “prompts.” No evidence buyers perceive or pay differently for the rebranding. | UNRESOLVED |
| 11 | Free GitHub norm won’t undercut | cursor.directory: 72.6K members, all free.6 SkillsMP: 200K+ skills from GitHub, all free.3 Open source culture is entrenched in dev tools. | CONFIRMED THREAT |
| 12 | Cursor won’t absorb this | Cursor’s marketplace is 2 weeks old. Private team marketplaces “coming soon.”8 They can add ratings, paid plugins, and featured placement at any time. | HIGH RISK |
Score: 4 confirmed threats, 3 weakened, 1 unresolved, 0 holding. The V1 stance has 0 assumptions that survived intact for static skill files. The only assumption that holds is #8 for functional tools (Apify model), not for taste/guidance skills.
[INVERSION] What if “taste” is NOT scarce? CONFIRMED — AI default output quality improves every model generation. The gap between “default” and “skilled” narrows structurally. Claude 4.6 and GPT-5.2 produce outputs that would have required a custom skill 6 months ago. The half-life of a “taste skill” is 3–6 months before the model catches up.
[BASE RATE] What’s the base rate for template/prompt marketplace success? CONFIRMED — 76% of ThemeForest themes earn <$1K/mo.19 GPT Store typical: $100–500/mo.18 80% of Shopify app merchants churn.13 The base rate for individual creators in digital asset marketplaces is <$1K/mo median across every platform studied.
[SURVIVORSHIP] Are we reasoning from winners? CONFIRMED — V1 cited Figma (12K plugins), Notion (fee schedule), Webflow (95% creator share) as evidence. But: Figma plugins are overwhelmingly free. Notion template top creators earn $80K total (not per year). Webflow templates are inside a $4B platform. The “proof of willingness to pay” evidence was cherry-picked from platform-owned marketplaces.
[INCENTIVE MAPPING] Who benefits from us believing “AEO/GEO is a budget line”? CONFIRMED — Conductor (the survey publisher) sells AEO/GEO tools. Their survey samples enterprise CMOs who already run AEO/GEO programs. Gartner’s independent data: only 23% of enterprises have optimized for AI discoverability.16 The 97% “measurable impact” claim is from people who are paid to make AEO/GEO work.
[TIME HORIZON] Is the ecosystem stable enough to build on? WEAKENED — Cursor Marketplace launched 2 weeks ago. MCP Registry is months old. SkillsMP grew 18x in 3 months. This is pre-infrastructure — building a marketplace on top of platforms that are still forming their own marketplace strategies is building on sand. But: early movers in Shopify apps (2012) and ThemeForest (2008) captured disproportionate value. The timing risk is real but not disqualifying.
UNCERTAIN — right demand signal, wrong product shape.
The demand for agent skill discovery is real and growing explosively (SkillsMP 18x in 3 months, 958K visits/mo). But every platform that achieved scale is free. The only marketplace paying creators meaningful money (Apify $500K+/mo) sells functional tools, not taste guidance. The price floor for static prompt/skill files is $1–7 (PromptBase), which doesn’t support a solo operator.
What V1 got wrong: Framing the product as “AEO-native taste curation.” AEO/GEO is real but tiny (~1% of web traffic) and the 97% Conductor figure is vendor self-serving. “Taste” is a depreciating asset — AI models improve every release. Free supply (200K+ skills) is overwhelming. IDE extension culture (from VS Code) expects free.
What V1 got right: The insight that most users are bad at steering agents is legitimate. The Cursor Marketplace + MCP Registry creating an ecosystem is real. The distinction between “prompts” (fragile, copyable) and “workflow IP” (eval harnesses, tool wiring) is directionally correct — but the market doesn’t yet recognize or pay for the difference.
The better product shape: Eric’s real asset is outcomes, not text files. His research reports command consulting-level value (HK$20K+/mo from Sourcy and Talent Coop). His skills are the engine behind those outcomes, not the product themselves. The working monetization is: (a) sell configured agent outcomes as a service (what he already does), (b) publish functional MCP tools on Apify (proven revenue model), or (c) wait for Cursor to add paid plugin support and be first with quality plugins. None of these require building a marketplace.
If insisting on marketplace path: Build 5 functional MCP tools (not taste skills), list on Apify (proven $500K/mo payout infrastructure), sell packs on Gumroad ($29–49), and list free versions on cursor.directory for distribution. Kill if 0–1 paid sales in 2 weeks. Don’t call it a “taste marketplace” — call it what Apify calls it: “build and monetize MCP servers.”
crm/state/user.json, .cursor/skills/. 24+ skills built, $3,400/mo Cursor spend, 68 Dunbar contacts.