Ranking x402/MPP endpoint categories RDCO could actually monetize
The question
Of the five candidate endpoint categories agents can pay for over x402/MPP — real-time market data, structured-research APIs, compute-on-demand, agent-memory-as-a-service, and dataset-licensing — which would be most monetizable for RDCO? Rank by latent agent demand, setup cost, and fit with RDCO's existing data assets, then name the top 1-2 to pursue. The agent-payments cluster (Tony Dang credential-brokering, Nick Prince x402 IC memo, MPP+Tempo proposal) established that RDCO can take agent payments; it never answered what to sell.
What we already know (from the vault)
- The buyer side is already built and proven. [[08-tooling/2026-05-01-mpp-tempo-integration-proposal]] provisioned a Tempo mainnet wallet, a capped/audited wrapper script, and verified a sub-second MPP round-trip. Critically, Phase 3 of that proposal explicitly defers the seller side ("micro-paywalled audit reports / model audits sold to other agents") — and flags Cloudflare Workers MPP middleware makes seller-side "trivial" since MAC/Squarely are already on Workers. This brief is the deferred Phase-3 question.
- The regulatory wrapper is clear: [[06-reference/research/2026-05-25-fincen-bsa-llc-x402-usdc-micropayments]] confirms an LLC selling its own data/research/compute for USDC is a "user," not an MSB — no registration, no inflow threshold. So the choice is unconstrained by compliance; pick on economics.
- The buyer-shape evidence points at instrumentation-sold-to-agents. [[06-reference/2026-05-01-alpha-vantage-collison-agent-as-customer-evidence]] documents Claude Code autonomously buying an Alpha Vantage dataset mid-report via MPP/Tempo — but the same note's RDCO-positioning warns instrumentation commoditizes fast ("10 competitors with MCP servers by year-end"); the defensible layer is targeting (which dataset answers which question), i.e. the MAC frame, not raw feeds.
- The concept is named but unranked. [[06-reference/concepts/products-for-agents]] lists the candidate surfaces (the QMD-indexed vault, Data Dots, Data Marketplace, MAC) and the three requirements — structured content, semantic context, composability — but leaves "how do you price a product-for-agents" and "which format" as open questions.
- The pitch motion already assumes seller-side. [[06-reference/2026-05-23-nick-prince-spacex-ipo-agent-ic-memo-x402]] reframes RDCO's sales motion around "fund this wallet, the agent buys its own data" — the IC memo cost $1.87 across 6 x402 calls. That is the demand-side shape RDCO would be selling into.
What the web says
- Aggregate x402 demand is real but small and heavily inflated. By late April 2026 x402 showed ~69K active agents, ~165M cumulative transactions,
$50M cumulative volume ($0.30/call avg). But CoinDesk reports roughly half of transactions are wash/self-dealing (seller funds buyer's wallet) and "the merchants x402 is designed to serve are still rare" — narrative ahead of usage (CoinDesk). - The single most decisive finding: skills convert ~33%; live data feeds convert 0-0.8% at identical price. A 368-probe field study found agent-testing/eval skills at 38%, agent-memory-context skills at 33%, while DeFi-yields-live (0.8%), token-anomalies-live (0%) and security-intel-live (0%) flatlined. The gap is informational, not economic — agents pay confidently for bounded, predictable, self-evaluable outputs and refuse to pay speculatively for open-ended "freshness." Adding a
data_freshness_secondsfield lowered conversion (dev.to / nathanielc85523). - That same study reframes the five categories: "Skills are task-bound, autonomous, discrete-output and fit the x402 model well, whereas data feeds may need a different monetization model such as subscription via MPP" (dev.to). This maps the x402-vs-MPP split onto endpoint type: discrete-output → x402 per-call; continuous feeds → MPP session/subscription.
- Counter-signal from the supply side: AWS argues per-query market/bureau/sanctions/regulatory data is the prime FSI monetization wedge — trading agents need real-time feeds, decisioning agents need bureau data, compliance agents need sanctions lists, all blocked today by subscription-only contracts (AWS). Note this is a latent-demand claim, not a realized-conversion claim — and AWS sells the infra, so it is supply-incentivized.
- Pricing reference points: a financial data point ~$0.001/call, a lead-enrichment record ~$0.02, the Nick Prince IC memo ~$0.10-0.30/call. Compute/skill invocations (translation, DB query, image gen, eval) are the named MCP-monetization sweet spot across Alchemy, Zuplo, and Nevermined writeups (Alchemy, WorkOS).
- Dataset-licensing has a real but narrow lane: agents buy "training data, market research, or proprietary datasets in real time without negotiating contracts" — the Alpha Vantage pattern — but this is bounded-output (a dataset is discrete), so it behaves more like a skill than a feed (Allium).
Convergences and contradictions
- Sharpest contradiction — and it's the whole answer. AWS (supply-side, financial-services-infra-incentivized) says real-time/per-query data feeds are the prime wedge. The 368-probe field study (demand-side, empirical) says data feeds convert at ~0% while skills convert at 33%. These disagree because AWS measures latent demand (the friction is real) and the probe study measures realized conversion (agents don't yet have autonomous budget for open-ended freshness, and can't self-evaluate a feed's value pre-purchase). For a solo operator picking what to ship in 2026, realized conversion is the binding constraint. Resolution: weight the empirical conversion data over the supply-side latent-demand pitch.
- Convergence: every source agrees the bounded/discrete-output property is what makes an endpoint payable. Skills, eval/audit runs, agent-memory-context lookups, and discrete dataset pulls all share it. Streaming freshness-dependent feeds lack it. This is consistent with the vault's own products-for-agents requirements (structured, semantic, composable) and with the Alpha Vantage note's warning that raw instrumentation commoditizes while the targeting judgment stays sticky.
- Convergence on rails-by-type: discrete output → x402 per-call; continuous subscription-shaped consumption → MPP sessions. RDCO's installed buyer stack already speaks both, and the FinCEN brief removes the entity-wrapper blocker for either.
- Caution flag: aggregate x402 volume is half-fake and merchant-thin (CoinDesk). This is an early-and-shallow-market signal, not a no-market signal — it argues for a low-setup-cost entry that doubles as phData demo material, not a heavy build betting on 2026 transaction revenue.
Synthesis for RDCO
Ranking the five categories by (latent agent demand × RDCO existing-asset fit) ÷ setup cost, using realized-conversion evidence as the demand weight:
- Compute-on-demand, narrowed to "skills as paid endpoints" — specifically MAC's data-quality audit / Scope×Basis test-generation run. Highest score. Demand: skills are the only category with proven ~33% conversion. Asset fit: RDCO already owns the targeting IP (
/audit-model,/generate-tests, the MAC matrix) — exactly the bounded, self-evaluable, discrete-output shape agents pay for, and exactly the "targeting not instrumentation" layer the Alpha Vantage note says is defensible. Setup cost: near-zero — Cloudflare Workers MPP middleware over an already-Workers-hosted surface, the deferred Phase 3 the proposal already scoped as "trivial." - Agent-memory-as-a-service, narrowed to the QMD-indexed vault as a queryable knowledge endpoint. Strong second. Demand: agent-memory-context skills clocked 33% in the probe study — the second-highest category. Asset fit: RDCO already runs this internally (vault + QMD semantic search + wikilinks is a product-for-agents built for ourselves, per the concept note); exposing a paid query endpoint is repackaging, not building. Setup cost: low-moderate (need scoping/redaction of what's queryable, plus a per-query meter). The risk is content sensitivity, not engineering.
- Dataset-licensing (discrete pulls), e.g. Data Dots / a curated data-engineering knowledge pack. Mid. Behaves like a skill (bounded output → payable), Alpha-Vantage-shaped demand is real, and Data Dots is already conceived as agent-ingestible. But RDCO's datasets are small and undifferentiated versus the commoditization warning; defensibility is thin.
- Structured-research APIs (e.g. a Sanity-Check-grade research-brief endpoint). Lower. This is really category 1 wearing a research hat — and the most defensible version is "agent pays for a bounded analysis run," which collapses back into skills. As a raw research-data API it inherits the feed problem.
- Real-time market data. Last — do not pursue. Despite AWS's pitch, this is the 0%-conversion category, RDCO has no proprietary market-data asset (it buys Alpha Vantage, doesn't produce a feed), and setup cost to be credible is high. This is the one to explicitly rule out.
Pursue #1 and #2. The move is to expose MAC's audit/test-generation as a per-call x402 skill endpoint (primary) and the QMD-vault as a metered agent-memory query endpoint (secondary), both via Cloudflare Workers MPP middleware on surfaces RDCO already operates. This is the lowest-setup-cost path, lands squarely on the only two categories with proven agent willingness-to-pay, and reuses assets RDCO already owns rather than building a feed it would have to source. Just as important, it is the strongest phData demo: "I built an agent that audits your dbt models and charges $0.10 per run, no seat license" is the sharpened sales motion the Nick Prince note already gestures at — making the build dual-purpose (small direct revenue + consulting credibility) and de-risking the thin/half-fake current x402 market. The honest caveat: 2026 x402 demand is shallow and partly wash-traded, so treat near-term revenue as secondary to the demo/learning value, and ship the lowest-cost version first.
Open follow-ups
- What is the minimal viable seller-side build for a MAC audit x402 endpoint on Cloudflare Workers MPP middleware, and how does it interact with the existing capped/audited buyer wrapper?
- For the QMD-vault-as-memory endpoint: what redaction/scoping boundary makes vault content safely queryable by external agents without leaking RDCO-internal strategy?
- Does the x402-vs-MPP rail split (discrete→x402, subscription→MPP) imply RDCO should offer both a per-call and a session-priced tier for the same MAC endpoint?
- What does a self-evaluable metadata response look like for a MAC audit endpoint (input schema + bounded-output description + price) — the exact thing the 368-probe study found drives skill conversion?
- Is there a Squarely-side agent endpoint (puzzle generation/validation as a paid skill) worth a parallel low-cost experiment?
Sources
- CoinDesk — Coinbase-backed x402 demand "just not there yet" (Mar 2026)
- dev.to / nathanielc85523 — Why AI Agents Pay for Skills but Not Data Feeds: 368-Probe Analysis
- AWS — x402 and Agentic Commerce in Financial Services
- Alchemy — How x402 brings real-time crypto payments to the web
- WorkOS — x402 vs Stripe MPP: choosing payment infra for AI agents and MCP tools in 2026
- Allium — x402 explained: payments for APIs, data, and agent commerce
- Vault: [[08-tooling/2026-05-01-mpp-tempo-integration-proposal]], [[06-reference/research/2026-05-25-fincen-bsa-llc-x402-usdc-micropayments]], [[06-reference/2026-05-01-alpha-vantage-collison-agent-as-customer-evidence]], [[06-reference/concepts/products-for-agents]], [[06-reference/2026-05-23-nick-prince-spacex-ipo-agent-ic-memo-x402]]