About
jpcite is a REST and MCP API that exposes Japanese public-program data — subsidies, loans, tax incentives, and certifications — to AI agents in a single call. 14,472 programs (11,601 searchable in tier S/A/B/C; 2,871 excluded from public search), plus 2,286 case studies (採択事例), 108 loan offerings, and 1,185 enforcement cases, are normalized from primary sources (METI, MAFF, SME Agency, Japan Finance Corporation), scored into tiers (S/A/B/C/X), cross-linked with 181 exclusion / prerequisite rules, searchable via a Japanese full-text search index (3-gram), and major public rows carry source_url and fetched_at for lineage.
Thesis — Sonnet/Haiku + jpcite produce Opus-4.7-grade Japanese-regulation outputs at up to ~80% lower cost (design estimate; Opus baseline = 7-turn Deep++ chain)
Opus 4.7 produces high-precision Japanese-regulation outputs because it reads long PDFs and multiple ministry pages in a 7-turn chain and structures them. jpcite does that "pre-fetch + structure + exclusion-rule resolution" ahead of time from its own corpus via deterministic composition, returning source_url, source_fetched_at, known_gaps, compatibility / exclusion rules, and citation candidates as a small Evidence Packet. A smaller model (Claude Sonnet 4.6 / Haiku, GPT-4o-mini, Gemini Flash) reads the packet once, so it processes the same decision material as an Opus 4.7 chain within a 1/17 – 1/167 cost envelope — by design (figures are design estimates, Opus baseline = 7-turn Deep++ chain; saving vs single-turn Opus is the ~3x band; to be updated LIVE after multi-customer empirical telemetry). jpcite does not generate the answer text — it is the Japanese-regulation packet bank that AI agents find and buy. The anonymous tier needs no signup (3 requests/day per IP); continued use is ¥3/billable unit (¥3.30 tax incl.). Verify with the cost-saving worked examples and the JCRB-v1 benchmark (5 cohorts × 50 queries, Opus vs jpcite).
Why it exists
jGrants is Japan's application portal. It is not a discovery or compatibility layer. Asking "which subsidy matches this business, and which of them can be combined? has anyone like us actually been selected? which loan has no collateral and no guarantor? has any vendor been sanctioned?" still requires a human to browse ministry sites, download PDFs, cross-reference notice documents, and keep up with ad-hoc revisions. jpcite is that missing layer, packaged as an API that AI agents can actually call. MCP protocol version 2025-06-18; no SDK required.
Who built it
Operated by Bookyou Inc. Hosted in Tokyo and billed via Stripe Tax. Legal entity and qualified-invoice details are available in the Japanese legal disclosure.
Quick facts
| Programs indexed | 14,472 (11,601 searchable in tier S/A/B/C; 2,871 excluded from public search) |
|---|---|
| Case studies (採択事例) | 2,286 |
| Loan programs (detailed) | 108 (with three-axis collateral / guarantor decomposition) |
| Enforcement cases | 1,185 |
| Laws (e-Gov CC-BY) | e-Gov law metadata and article references; coverage varies by record. |
| Tax rulesets | 50 |
| Court decisions | 2,065 |
| Bids | 362 |
| Invoice registrants | 13,801 (PDL v1.0 delta-only; full ~4M monthly bulk pending) |
| Entity-fact DB entities | 503,930 (across 12 record_kinds) |
| Entity-fact DB facts | 6.12M |
| Entity-fact DB relations | 177,381 (15 canonical relation types) |
| Primary-source coverage | Most public rows carry source_url + fetched_at; gaps are flagged instead of backfilled |
| Exclusion rules | 181 (125 exclude + 17 prerequisite + 15 absolute + 24 other) |
| MCP tools | 261 at standard configuration, protocol 2025-06-18 (program / case-study / loan / enforcement search, source-backed context packets, metadata tools, and DD-support tools) |
| MCP protocol | 2025-06-18 |
| Region | Japan |
| Launch | 2026-05-06 |
Contact
Email: [email protected]. For press, the full kit is at /press/about.md. For Japanese legal disclosures (特定商取引法), see the JP tokushoho page.