MCP server charges AI agents per call in USDC for Korean government data

A developer in Seoul has published a working demonstration of what agent-native commerce might look like: a remote MCP (Model Context Protocol) server that answers questions about Korean businesses using official National Tax Service data and charges AI agents per call in USDC, with no signup, API key or human intervention. The write-up on dev.to describes the project, Korea Business Verify (KBV), as a first end-to-end test of the model.

A government API rebuilt for machine clients

According to the author, the underlying data is free — Korea's tax authority operates a public API that reports whether a business registration number is active, suspended or closed — but it sits behind a Korean-language portal, phone-based signup and manual key issuance. An autonomous agent cannot pass SMS verification or read the documentation, so the official source is effectively unusable for non-human clients. KBV wraps that API in a single URL that returns status, tax type and identity-matching results as English JSON, with batch lookups of up to 100 companies per request. The code is open source under an MIT licence and runs on Cloud Run with scale-to-zero, so idle hosting cost is close to zero; the developer puts total build and deployment spend at roughly $3.

Payments over HTTP 402

Charging works through x402, a payment protocol built around the HTTP 402 status code. Per the dev.to post, pricing is $0.02 for a status check, $0.05 for an identity verification and $0.02 per company in batch mode, settled in USDC on Base via a Coinbase CDP facilitator. The server itself never holds a private key — it only knows a receiving address, and payment verification is delegated to the facilitator, with the developer's keys kept in a browser wallet. Notably, this works even though Coinbase's retail exchange is unavailable in Korea, because receiving payments requires only an address. The author reports that the first mainnet payment succeeded end to end: an agent hit a 402, paid two cents and received real government data back.

Directory traffic was mostly bots

After listing the server on the official MCP Registry, Glama and Smithery, it began drawing about 1,250 requests a day. A user-agent breakdown showed nearly all of that volume came from health checks and security scanners — some openly stating in their user-agent strings that they never invoke tools. Genuine external tool calls in the first week were effectively zero. The author's takeaway: request counts are a poor popularity signal for MCP servers, and builders should count actual tool invocations instead.

Free tiers and paid discovery conflict

One of the more consequential findings concerns x402's discovery marketplace, the Bazaar, whose validator checks that an unpaid request returns HTTP 402. KBV's free tier — ten calls per day per IP address — returned a 200 instead, so the validator judged the resource unpayable and excluded it, while also letting scrapers rotate addresses for unlimited free access. The fix was to make the REST endpoint return 402 by default and turn the free tier into an opt-in query parameter. The author argues that paid-by-default, free-only-on-request is the better posture for services aimed at agents.

Everything drifts

The post also catalogues ecosystem churn: the MCP registry rejected the first two publish attempts because the server. schema had changed weeks earlier; x402 has moved to a v2 API while most tutorials still cover v1; and a fresh Cloud Run URL took four hours to propagate. Separately, a trust-scoring service called M8ven automatically indexed the server, ran static analysis that found no credential exfiltration or obfuscation, and flagged missing MCP annotation hints (such as readOnlyHint) required by OpenAI's directory — which the author fixed the same day. In their view, the agent ecosystem is building its defenses faster than its customer base.

Why it matters

KBV is less a business than a proof, but a meaningful one. It demonstrates that per-call, machine-to-machine payment for data services can work end to end today, using infrastructure — MCP for access, x402 for billing — that any developer can adopt. The monetized value is not the data itself, which the government gives away, but the removal of friction that excludes automated clients. The post is equally useful as a reality check for anyone building in this space: early traffic will be bots rather than buyers, generous free tiers can disqualify you from paid marketplaces, and documentation will lag the protocols it describes. The author describes the pool of paying agents as real but early, and frames the project as claiming a niche — official Korean data, payable by agents, zero setup — before automated procurement becomes routine.