how it works
How jpcite Evidence Packets work for AI agents
Your AI receives a Japan-regulatory question. It hits the jpcite Evidence Packet bank, retrieves a structured JSON envelope (source URL, fetched-at, known gaps, Ed25519 attestation), then answers using verified evidence instead of long-context retrieval. Two paths: browser AI uses Apple Pay 1-tap, CLI AI uses MCP integration.
Browse the catalogue See pricingThree-step funnel
- Discover — Browse the catalogue at /en/packets/ or call
jpcite.list_packets()from your AI agent over MCP. - Verify — Free anonymous evaluation gives 3 requests/day per IP (no signup, resets at JST next-day 00:00). The returned envelope contains
source_url,source_fetched_at,known_gaps, exclusion/prerequisite rules, and an Ed25519 attestation. - Integrate and pay — Pay-as-you-go ¥3 per billable unit. Stripe metered billing (qualified Japanese invoice issued automatically) or x402 USDC native payment over HTTP 402 (no API key required, <2 sec settlement).
Two integration paths
Browser AI (Claude.ai / ChatGPT / Perplexity)
Your agent calls the REST endpoint and surfaces an Apple Pay 1-tap purchase link. The Evidence Packet is unlocked within seconds and embedded into the answer.
CLI AI (Cursor / Claude Code / Windsurf)
Install the MCP server: npx @bookyou/jpcite or uvx autonomath-mcp. The agent calls jpcite tools natively over MCP stdio. Stripe metered billing or x402 wallet auto-tops-up.
What's in an Evidence Packet
| Field | Meaning |
|---|---|
source_url | Primary public-source URL (METI / MAFF / NTA / Courts / e-Stat / etc.) |
source_fetched_at | ISO-8601 timestamp when jpcite last crawled the source (≤24 h for hot packets) |
known_gaps | Explicit known limitations (12 small-municipality programs lack a dedicated CMS page) |
exclusion_rules | 181 codified exclusion + prerequisite rules (125 exclude + 17 prerequisite + 15 absolute + 24 other) |
citation_candidates | Suggested citation form for the LLM answer step |
attestation | Ed25519 signature over the packet JSON for tamper-evidence |
Why agents save tokens with jpcite
Thesis: Sonnet 4.6 / Haiku / GPT-4o-mini / Gemini Flash + jpcite Evidence Packet ≈ Opus 4.7 7-turn over raw Japanese government PDFs, at 1/17 – 1/167 of the LLM token cost. The structured ground truth Opus would re-derive each turn is already in the packet.