Mart402 is a set of pay-per-call APIs built for AI agents, in three lines: documents & web (web-to-Markdown, PDF parsing, OCR with hallucination detection, invoice verification, company enrichment), Japan statutory calculations (payroll, social insurance, tax - deterministic and primary-source cited) and calibrated vision (meal nutrition, personal color). There is no account and no API key; agents pay per request with x402 (USDC on Base). Prices start at $0.001 per call. See /catalog for machine-readable pricing.
Mart402 uses x402, the HTTP 402 payment protocol: the first request returns HTTP 402 with machine-readable payment terms (price, asset, recipient), the agent signs a USDC payment authorization and retries with an X-PAYMENT header, and the result comes back in the same HTTP exchange. No signup, no subscription, no key management. Payment examples: https://github.com/tanaka-77/mart402-agent-kit
OCR engines fail silently: a vision-language model can output fluent text that is not what the document says. Mart402 runs a second, independent engine with orthogonal failure modes, cross-checks the two reads, and returns a consensus verdict with calibrated confidence and per-segment corroboration - so an agent that cannot taste-test the output can still act on it mechanically. Details: /ocr
Yes - Japanese-strong OCR routing is a core design goal: dual-engine consensus tuned for Japanese text, script-mismatch detection (e.g. Devanagari appearing in a Japanese document), and invoice field checks for Japanese accounting formats. OCR is a flat $0.05 per image (dual-engine consensus); PDFs are $0.01 per document when text-embedded and $0.015 per page when scanned.
Yes, three ways: a sandbox with the same API on Base Sepolia testnet (https://mart402.dev, free test USDC), free PDF quotes on POST /v1/parse/quote, and recorded sample responses for the LLM-backed products on the sandbox. Calling any paid endpoint without payment safely returns the exact terms instead of charging.
No. Since 2026-09-27 every product verifies the payment before running and settles it on-chain only after the call succeeds, so a failed call is never charged (the response carries billed_usd 0). Errors are structured ({code, retryable, action}) and say whether retrying makes sense. If the on-chain settlement itself fails after a successful call, the result is held for one hour and repeating the request with a fresh payment delivers it without re-execution. Details: /terms (v1.4) and /catalog.billing.
Nothing is lost: resending the same payment returns the stored result without a second charge (idempotency), and every paid response carries a receipt id that re-fetches the result for free, forever within retention, at GET /v1/receipts/{id}.
Yes - https://mart402.com/mcp is a hosted MCP endpoint (Streamable HTTP) that exposes every product as a tool, plus free tools (catalog, PDF quote, receipt re-fetch, samples, x402 service search) that work immediately; paid tools called without payment return the exact payment terms instead of charging.
Yes - this is a common agent workflow: before paying an unfamiliar x402 seller, check its measured record on Market402 (https://market402.com/op/{operator} - score, history, verification state); after ordering, extract the order-confirmation page (/v1/extract), parse the PDF receipt or invoice (/v1/parse), and verify the amounts add up (/v1/verify). Every step is pay-per-call with its own receipt.
Do not take our word for it: Market402 (https://market402.com/op/mart402.com) probes Mart402 with real requests under the same published rules it applies to 1,000+ other x402 operators. Market402 is run by the same operator, so its rules are public (market402.com/methodology) and every paid claim carries an on-chain tx hash you can verify yourself. Measured latency and reliability are also public at /stats.
Yes. POST /v1/meal-vision turns one meal photo into itemized nutrition JSON (dishes, portions, kcal, protein/fat/carb) for $0.05. Every call is a 3-run consensus that discloses its own kcal spread and item agreement, and the response includes a calibrated two-sided kcal range instead of a single point guess. Workflow: /use-cases/meal-logging. Evidence: /vision-calibration.
Measured, not claimed: against 30 dietitian-portioned weighed meals (Japanese, Chinese, Western), the median error is about 12-15 percent and the disclosed kcal range contained the true value in 26 of 30 meals. Systematic error classes we found and corrected are published openly. It is not for allergen detection or medical decisions. Full method and findings: /vision-calibration.
Yes - measured, not assumed. We verified the full flow end-to-end with @x402/fetch v2.21.0, the standard x402 client the Cloudflare Agents SDK uses: the 402 challenge parsed, the EIP-3009 payment signed, settled and the result delivered (Base Sepolia sandbox; production speaks the identical x402 v2 protocol). Any agent using @x402/fetch or x402-hono tooling can pay Mart402 without custom integration. Try it on the sandbox: https://mart402.dev (test USDC from faucet.circle.com).
POST /v1/portrait-analysis measures skin colorimetry from 1-3 portraits (CIELAB, ITA angle - measured pixels, not model vibes), cross-checks it against a vision model, and applies a published rule table for a 4-season verdict with all season scores and a structured palette. $0.05 flat; a free precheck (/v1/portrait-precheck) tells you photo quality before you pay. No identification; images are not stored. Workflow: /use-cases/personal-color.