jpcite

jpcite

日本の公的制度(補助金/法令/税務/法人/判例)を提供する MCP サーバー。261 ツール、¥3/billable unit、匿名 3/日 free。Evidence Packets with source_url + source_fetched_at + known_gaps. PyPI: autonomath-mcp.

Category
Visit Server

README

<p align="center"> <a href="https://jpcite.com"><img src="https://jpcite.com/assets/github-social-card.png" alt="jpcite — Japan regulatory MCP server (60-second answers for AI agents)" width="800"></a> </p>

jpcite

packets = ingredients, your AI agent finishes

日本の制度・税法・法令の ingredient (素材) を AI agent に提供する MCP server.

agent が packet を読み、user 文脈で finishing (~500 token, ~¥10) して end-user に返す前提. 合計 ¥30-¥45 (ingredient + finishing) で LLM 単独 ¥120 より 67% 安い + 出典付き.

English: jpcite packets are ingredients (foundation data), not final answers. AI agents read packets and apply finishing (~500 tokens, ~¥10) to produce end-user deliverables. Total cost: ingredient (¥3-¥30) + finishing (¥5-¥15) = ¥30-¥45 typical, vs LLM-only ¥120.

<p align="center"> <a href="https://jpcite.com/try"><img src="https://img.shields.io/badge/Try%20free-3%20req%2Fday-4c1.svg?style=for-the-badge" alt="Try free"></a> <a href="https://pypi.org/project/autonomath-mcp/"><img src="https://img.shields.io/badge/pip%20install-autonomath--mcp-3776AB.svg?style=for-the-badge&logo=pypi&logoColor=white" alt="pip install autonomath-mcp"></a> <a href="https://github.com/shigetosidumeda-cyber/autonomath-mcp"><img src="https://img.shields.io/badge/GitHub-Star%20%E2%98%85-181717.svg?style=for-the-badge&logo=github&logoColor=white" alt="GitHub Star"></a> </p>

mcp-name: io.github.shigetosidumeda-cyber/autonomath-mcp

v0.4.0 LIVE on Fly.io Tokyo — production at api.jpcite.com. Current public docs, manifests, and release tags are the source of truth for version and pricing.

PyPI version License: MIT Python 3.11+ MCP 2025-06-18 Made in Japan

Smithery Glama mcp.so mcphub.dev awesome-mcp

mypy strict pytest production gate API status

<!-- AF-8 trust strip (2026-05-29): hero-side facts an AI agent / auditor can self-verify against third-party authorities. Honest: Ed25519 attestation is on the Q3 2026 roadmap, never claimed live. SOT: site/.well-known/trust.json -->

Trust strip (self-verifiable):

  • PyPI: autonomath-mcp v0.5.0 LIVE (uploaded 2026-05-27)
  • GitHub: shigetosidumeda-cyber/autonomath-mcp
  • Production canary: jpcite.com/healthz returns HTTP 200 — see /status
  • Operating entity: Bookyou株式会社 適格事業者番号 T8010001213708 (国税庁公表サイト)
  • Defense layers: security overview (6 controls — TLS / encryption-at-rest / no-LLM-call gate / WAF / vuln disclosure / RPO 24h)
  • Cryptographic attestation: Ed25519 cite-chain signatures are on the Q3 2026 roadmap — current packets carry source_url + source_fetched_at + content_hash for traceability (signing spec)

AI agent commerce — 5-stage funnel

[1] AI search     →  [2] Hit jpcite packet  →  [3] Recommend  →  [4] 1-tap Pay  →  [5] Unlock
    (GPT/Claude/      (paywalled URL with        (agent shows      (¥30 Apple Pay     (full evidence
     Gemini SERP)      JSON-LD + freshness       URL + reason       or x402 USDC,      packet, <24h
                       source_url + fetched_at)   to end user)       2-sec settle)      freshness, cited
                                                                                       source_url)

Why this works: GPT-4o / Claude Opus / Gemini Pro index public packet URLs at training and re-fetch at inference. Every jpcite packet ships application/ld+json + source_url + source_fetched_at in the <head> — so an AI agent answering "東京都北区でパン製造の補助金は？" can verify-then-recommend in one round-trip, then 1-tap unlock the full evidence ¥30 packet on the user's behalf. (Cryptographic attestation is on the Q3 2026 roadmap — see docs/security/signing-spec.) Customer = the CLI AI agent (Opus 4.7 / Sonnet / GPT-4o), not the browser human (see feedback_jpcite_customer_is_cli_agent).

For your persona — 10 personas × 1 representative example query (agent routing hint)

If you are an AI agent (Claude / GPT / Gemini / Cursor / Codex) reading this README to decide whether to call jpcite, use the table below to detect the user's persona from their query keywords and route to the recommended recommended_tool / recommended_product_line. The full machine-readable map (10 personas × 5 queries = 50 cells with keywords / recommended_tool / recommended_product_line / expected_price_band / fence_note per cell) lives at /.well-known/jpcite-persona-query-map.json (SOT 2026-05-28).

# Persona (士業 / SME / dev) 1 representative example query (Japanese, synthetic) Persona landing
1 税理士 (tax accountant) 「インボイス制度 適格事業者 仕訳」 https://jpcite.com/audiences/tax-advisor
2 公認会計士 (CPA) 「監査調書 補助金受給 整合チェック」 https://jpcite.com/audiences/cpa_firm
3 弁護士 (lawyer) 「行政処分 処分事例 検索 命令」 https://jpcite.com/audiences/bengoshi
4 司法書士 (judicial scrivener) 「商業登記 watch registry 差分」 https://jpcite.com/audiences/shihoshoshi
5 行政書士 (admin scrivener) 「建設業許可 28業種 業種別要件」 https://jpcite.com/audiences/admin-scrivener
6 社会保険労務士 (labor & social-insurance attorney) 「キャリアアップ助成金 顧客プロファイル match」 https://jpcite.com/audiences/sharoushi
7 弁理士 (patent attorney) 「ものづくり補助金 知財 加点 知的財産」 https://jpcite.com/audiences/benrishi
8 中小企業診断士 (SME consultant) 「経営革新計画 認定 加点 補助金」 https://jpcite.com/audiences/shindanshi
9 中小企業経営者 (SME owner) 「補助金 検索 採択率 業種別」 https://jpcite.com/audiences/smb
10 AX エンジニア (AI agent dev / SaaS backend) 「MCP install Claude Code Cursor uvx npx」 https://jpcite.com/audiences/dev

Claim-guard: queries above are representative examples (synthetic, for AI-agent routing only) — they are NOT a claim that real customers have sent the verbatim text. jpcite returns evidence-substrate ingredients (法令 raw text / 公的 program JSON / 適格事業者 row / 判例 PDF) only; it does not provide 税務助言 / 法律相談 / 監査結論 / 社労判断 / 知財判断 / 経営革新認定 — 8 業法 fence (税理士法 §52 / 弁護士法 §72 / 公認会計士法 / 司法書士法 §3 / 行政書士法 §1 / 社労士法 §27 / 弁理士法 §75 / 労働基準法 §36) is auto-injected in every response envelope. Final professional judgment stays with the licensed practitioner.

5 sample packets (AI-search SERP test corpus)

Drop these URLs into ChatGPT / Claude / Gemini and ask "is this evidence reliable?" — the agent will read the JSON-LD preamble + check source_url + source_fetched_at + cite back:

# Topic Packet URL
1 北区パン製造補助金 (SME) https://jpcite.com/packets/sme-kita-ku-pan-seizou-hojokin-2026-05-18.html
2 軽減税率対象品目 (税理士) https://jpcite.com/packets/zeirishi-keigen-zeiritsu-2026-2026-05-18.html
3 法務局 routing (司法書士) https://jpcite.com/packets/shihou-shihou-shoshi-vs-houmu-kyoku-routing-2026-05-18.html
4 商標出願戦略 (弁理士) https://jpcite.com/packets/benri-senkou-shouhyou-search-routine-2026-05-18.html
5 法人税試算 (税理士) https://jpcite.com/packets/zeirishi-bouei-tokubetsu-houjinzei-2026-2026-05-18.html

Full catalog: https://jpcite.com/packets/ (500+ packets, freshness <24h, source-linked, agent.json indexed).

Quickstart — Cursor / Claude Code (30 seconds)

Drop into ~/.cursor/mcp.json or ~/.claude/mcp.json:

{ "mcpServers": { "jpcite": { "command": "uvx", "args": ["autonomath-mcp"] } } }

Restart your client. Ask: 「東京都北区でパン製造の補助金を教えて」 — your agent will hit jpcite, pay ¥3-¥30 per call from your credit wallet, and cite back with source_url + freshness.

REST one-liner (no MCP client needed):

curl -G "https://api.jpcite.com/v1/programs/search" --data-urlencode "q=パン製造" --data-urlencode "prefecture=東京都" -H "X-API-Key: jc_xxx"

Opus alone vs Opus + jpcite

Dimension Opus 4.7 alone (7-turn chain) Opus 4.7 + jpcite (1-turn evidence)
Cost / query ¥347-¥500 ¥3-¥30 (jpcite tier A-D)
Freshness Training cutoff (months stale) <24h (source_fetched_at stamped)
Citation Hallucinated / outdated URLs Verified source_url + source_fetched_at (cryptographic attestation: Q3 2026 roadmap)
Japan regulatory recall ~60% on JCRB-v1 (scaffold) ~95% (design target) — empirical benchmark in progress, 5/250 fixtures populated (see P5 benchmark)
8業法 fence Manual disclaimer Auto-injected _disclaimer envelope
Settlement API key + monthly invoice 1-tap Apple Pay / x402 USDC

Per-call savings: 17x (Tier D) ... 167x (Tier A). Full math: docs/_internal/JPCITE_COST_ROI_SOT_2026_05_17.md.

Trust footer

  • Operator: Bookyou 株式会社 (T8010001213708) — 国税庁適格事業者番号
  • Attestation roadmap: Cryptographic cite-chain attestation (Ed25519 signature header + public key + verify endpoint) is on the Q3 2026 roadmap — see docs/security/signing-spec. Current responses carry source_url + source_fetched_at + content_hash for traceability; signatures are not yet emitted.
  • Freshness: source_fetched_at <24h on all hot tables; per-source distribution at /v1/stats/freshness
  • Compliance: 8業法 disclaimer envelope (税理士法 §52 / 弁護士法 §72 / 公認会計士法 §47の2 / 行政書士法 §1の2 / 司法書士法 §3 / 社会保険労務士法 §27 / 弁理士法 §75 / 労働基準法 §36)

TL;DR for AI agents

Thesis (canonical positioning, 2026-05-18): Sonnet 4.6 / Haiku / GPT-4o-mini / Gemini Flash + jpcite agent = Opus 4.7-equivalent Japan regulatory deliverables at 80% cost reduction. Down-model agents read jpcite Evidence Packets (pre-structured by 504K canonical entities × 6.2M source-backed facts + 181 exclusion/prerequisite rules + 8 業法 fence) instead of running a 7-turn Opus 4.7 chain themselves. The packet contains the structured ground truth Opus would re-derive turn-by-turn, so a down-model can match Opus output quality for Japan regulatory tasks (税理士月次 / 監査調書 / 適格性チェック / 登記 watch / 補助金 lineage / DD / M&A) at ¥10-¥40 per call instead of ¥347-¥500.

What it is: 261-tool MCP / REST surface — the Japanese-regulation packet bank AI agents find and buy — over Japanese public-program data (補助金 / 法令 / 判例 / 税制 / 適格事業者 / 法人 360° / 行政処分 / 入札 / 融資) with source_url, content_hash, source_fetched_at on every record.

How to use: uvx autonomath-mcp for stdio MCP, or X-API-Key against https://api.jpcite.com/v1/* for REST. No OpenAI / Anthropic / Gemini key needed — jpcite calls no LLM in production; reasoning stays in the caller's agent. Bring your own small/cheap model — jpcite Evidence handles the regulatory accuracy floor.

Cost: ¥3 / billable unit (tax-excluded; ¥3.30 inc. tax). Tier A=¥3 / B=¥6 / C=¥12 / D=¥30 — every tier is 1/17x to 1/167x of a Claude Opus 4.7 7-turn evidence chain (~¥500). Anonymous 3 req/day free per IP, JST reset. Empirical equivalence benchmark in progress (scaffold, 5/250 fixtures populated): see docs/_internal/P5_BENCHMARK_RESULTS_2026_05_17.md (JCRB-v1 50-query Opus vs Sonnet+jpcite).

Tier pricing — Opus 4.7 anchor comparison

Single billable-unit is ¥3; "tier" is the narrative bundle depth label tied to the agent-economy band. Full derivation: docs/_internal/JPCITE_COST_ROI_SOT_2026_05_17.md §3 (FF1 SOT, validator-gated).

Tier jpcite ¥/req Opus equiv turns Opus equiv ¥ Saving % Saving ¥ Default tool families
A ¥3 3 (light) ¥54 94.4% ¥51 search_*, list_*, get_simple_*, enum_*
B ¥6 5 (medium) ¥170 96.5% ¥164 search_v2_*, expand_*, get_with_relations_*
C ¥12 7 (deep) ¥347 96.5% ¥335 HE-1, HE-3, precomputed_answer, agent_briefing, cohort_*
D ¥30 7 (deep+) ¥500 94.0% ¥470 HE-1 full, evidence_packet_full, portfolio_analysis, regulatory_impact_chain

Saving ratio envelope: min 17x (Tier D) ... max 167x (Tier A) vs Opus 4.7 7-turn Deep++ tool-calling chain @ ¥150/USD FX. Per-case API-fee-delta form only. Aggregate profit/return forecasts are excluded — we do not claim revenue or earnings outcomes (see docs/canonical/cost_saving_examples.md).

How it works (ingredient + finishing)

jpcite is the ingredient supplier, not the chef. The AI agent in your CLI / IDE / Custom GPT does the finishing — composing the final, user-shaped answer from the structured packet jpcite ships. End-user never sees raw packets.

6-step flow (end-user → answer):

[1] end-user asks       →  [2] AI agent receives    →  [3] agent discovers jpcite
    "東京都北区で                 query in caller             via SERP / agent.json /
    パン製造の補助金は？"          context (chat / IDE /        Smithery / .well-known
                                  ticket / mail thread)        and selects 1 tool

[4] agent fetches       →  [5] agent finishes       →  [6] end-user reads
    ingredient packet           in user context             tailored answer
    (¥3-¥30, <500ms,            (~500 token, ~¥10           with cited 補助金 list
    source_url + fetched_at,     LLM cost) — tone /          + amount + deadline
    JSON-LD + known_gaps)        format / language /         + source links
                                 user-history weave

Cost breakdown (single end-user query):

Component Who pays Cost What it buys
ingredient (jpcite packet) agent → jpcite ¥3-¥30 structured, source-linked, <24h-fresh regulatory facts
finishing (LLM finishing turn) agent → LLM provider ~¥10 user-context shaping, language, tone, ~500 token gen
Total end-user (via agent margin) ¥30-¥45 one delivered answer with citations
LLM-only baseline (no jpcite) agent → LLM provider ~¥120 7-turn Opus 4.7 chain, no citations, training-cutoff stale

Net delta: ¥75-¥90 saved per query (~67%) plus citations (source_url + source_fetched_at + known_gaps) the LLM-only path can't produce. (Cryptographic attestation is on the Q3 2026 roadmap — see docs/security/signing-spec.)

5 cohort scenarios (full table at docs/canonical/cost_saving_examples.md):

  • 税理士 月次 — Tier B ingredient ¥6 + finishing ¥10 = ¥16/query vs LLM-only ¥120 (87% saving, audit-trail citations).
  • 会計士 監査調書 — Tier C ¥12 + finishing ¥10 = ¥22/query vs ¥347 (94% saving, 出典 lineage).
  • 行政書士 適格性 — Tier B ¥6 + finishing ¥10 = ¥16/query vs ¥170 (91% saving, 8業法 fence auto-injected).
  • 司法書士 登記 watch — Tier A ¥3 + finishing ¥5 = ¥8/query vs ¥54 (85% saving, 30 watch/月 batch).
  • SME / 補助金 — Tier C+D ¥30 + finishing ¥15 = ¥45/query vs ¥500 (91% saving, lineage + acceptance probability).

The agent is the chef. jpcite ships the ingredients. End-user gets a plate.

MCP server quickstart (Claude Desktop, 30 seconds)

{
  "mcpServers": {
    "jpcite": {
      "command": "uvx",
      "args": ["autonomath-mcp"],
      "env": {
        "JPCITE_API_KEY": "jc_xxx",
        "JPCITE_API_BASE": "https://api.jpcite.com"
      }
    }
  }
}

Drop into ~/Library/Application Support/Claude/claude_desktop_config.json, restart Claude Desktop, then ask: 「東京都で設備投資に使える補助金を教えて」. The JPCITE_API_KEY is for jpcite metered billing — not an LLM provider key. uvx-installed wheels ship without DB and auto-fall back to api.jpcite.com over HTTP for the top 10 tools (search_programs, get_program, search_case_studies, search_loan_programs, search_enforcement_cases, search_tax_incentives, search_certifications, list_open_programs, dd_profile_am, rule_engine_check). Other tools return error: "remote_only_via_REST_API" with the REST URL. Clone the repo for the full local-DB surface.

REST quickstart (30 seconds)

# Always --data-urlencode JA params — raw 補助金/設備投資 breaks curl's HTTP request line.
curl -G "https://api.jpcite.com/v1/programs/search" \
  --data-urlencode "q=設備投資" \
  --data-urlencode "prefecture=東京都" \
  -H "X-API-Key: jc_xxx"

Get a key: https://jpcite.com/pricing.html#api-paid. Dashboard / usage / billing: https://jpcite.com/dashboard.

Output sample

GET /v1/programs/search?q=設備投資&prefecture=東京都 (truncated to 1 result):

{
  "total": 47,
  "results": [
    {
      "unified_id": "UNI-example-energy-dx",
      "primary_name": "東京都 中小企業 省エネ設備導入支援",
      "amount_max_man_yen": 500,
      "application_window": {"end_date": "2026-06-30"},
      "source_url": "https://www.metro.tokyo.lg.jp/.../energy-dx.html",
      "source_fetched_at": "2026-04-30T00:00:00+09:00",
      "tier": "A"
    }
  ]
}

Cohort coverage — 5 cohorts × per-call saving

Mix-weighted annual API-fee-delta examples (per-cohort 100 query / year / user, mirrors site/pricing.html):

Cohort Tier mix jpcite ¥/yr Opus ¥/yr Saving ¥/yr Ratio
税理士 (tax-firm) 70 B + 30 C ¥780 ¥22,310 ¥21,530 28.6x
会計士 (CPA / audit) 40 B + 60 C ¥960 ¥27,620 ¥26,660 28.8x
行政書士 60 B + 40 C ¥840 ¥23,990 ¥23,150 28.6x
司法書士 (登記 watch) 60 A + 40 B ¥420 ¥10,040 ¥9,620 23.9x
SME / 補助金 30 B + 50 C + 20 D ¥1,380 ¥36,910 ¥35,530 26.7x

Per-product cohort packs (from FF1 SOT §4): A1 税理士 月次 12 packets/yr @ ¥6 → ¥72 vs ¥6,000 (83.3x). A2 会計士 監査 10 件 @ ¥12 → ¥120 vs ¥3,000 (25.0x). A3 行政書士 適格 1 件 @ ¥6 → ¥6 vs ¥170 (28.3x). A4 司法書士 登記 30 watch/月 @ ¥3 → ¥90 vs ¥1,620 (18.0x). A5 SME 補助金 5 候補 @ ¥12 → ¥60 vs ¥1,735 (28.9x).

Public copy uses "API fee delta" language only — see docs/canonical/cost_saving_examples.md. No return-multiple / labor-reduction / business-outcome claims (per feedback_cost_saving_not_roi guard).

Data moat — live corpus snapshot

Source-linked records carry source_url + content_hash + source_fetched_at lineage; known gaps surfaced explicitly. Aggregator pages are excluded from citation sources where detected.

Surface Live count Note
Canonical entities (am_entities) 504,238 法人 / 制度 / 法令 / 判例 / 採択 / 行政処分 unified ID space
Entity facts (am_entity_facts) 6,228,893 source_id-backed atomic facts (A6 done, source_id 0→81,787 backfilled)
Precomputed answers (am_precomputed_answer) 5,473 cohort × question Tier C/D bundles
Searchable programs 11,601 47 prefectures + national; tier S=114 / A=1,340 / B=4,186 / C=5,961
Full program catalog 14,472 + 2,871 publication-review rows
Laws full-text indexed 6,493 e-Gov CC-BY (out of 9,484 metadata stubs)
Tax rulesets 50 structured 措置法 + 通達 cross-ref
Invoice registrants 13,801 国税庁 適格事業者 PDL v1.0 delta
採択事例 2,286 + 108 融資 (担保/個人保証人/第三者保証人 三軸)
行政処分 1,185 + 22,258 enforcement-detail rows
Court decisions 2,065 + 362 bids
Exclusion / prerequisite rules 181 125 exclude + 17 prerequisite + 15 absolute + 24 other

agents.json corpus snapshot (2026-05-07) shows 503,930 entities / 6.12M facts at snapshot time; live values trump snapshot during drift windows.

Verify links (agent-readable SOT)

Surface URL Purpose
.well-known/agents.json https://jpcite.com/.well-known/agents.json machine-readable AI-capability spec (tools / pricing / corpus snapshot)
OpenAPI agent-safe https://api.jpcite.com/v1/openapi.agent.json ChatGPT Custom GPT Actions importer
OpenAPI full https://api.jpcite.com/v1/openapi.json SDK generators / Postman
Cost-saving SOT (FF1) docs/_internal/JPCITE_COST_ROI_SOT_2026_05_17.md tier quintuple (yen, opus_turns, opus_yen, saving_pct, saving_yen) — validator-gated
Cost-saving public copy docs/canonical/cost_saving_examples.md API-fee-delta narrative, 14 audience entries + 6 use-case calculator
MCP tool catalog docs/mcp-tools.md full 261-tool list + arguments
Distribution manifest scripts/distribution_manifest.yml canonical published counts (tool/route/openapi)
llms.txt (JA / EN) https://jpcite.com/llms.txt / https://jpcite.com/llms.en.txt AI-agent discovery surface
Benchmark (FF3 / P5) docs/_internal/P5_BENCHMARK_RESULTS_2026_05_17.md quality / latency benchmark walk
Evaluation suite evals/gold.yaml 79-query gold-standard (run .venv/bin/python evals/run.py)
Stats — coverage / freshness /v1/stats/coverage, /v1/stats/freshness, /v1/stats/usage live transparency endpoints

How jpcite compares to single-source MCP servers

jpcite is the 横断 + Evidence Packet layer. The 3 active single-source Japanese MCP servers each handle one slice — they are complementary, not competitive:

  • vs jgrants-mcp (digital-go-jp/jgrants-mcp-server, 5 tools, jGrants 補助金 only): jpcite adds 法令 / 判例 / 行政処分 / 適格事業者 / 法人 360° / 排他併用判定. Use jgrants-mcp for the grant application path; use jpcite for cross-source compliance check. → /compare/jgrants-mcp/
  • vs tax-law-mcp (kentaroajisaka/tax-law-mcp, 7 tools, e-Gov + NTA + KFS live scrape): jpcite adds 50 structured tax_rulesets + 9,484 e-Gov laws + 28,201 article rows pre-indexed (median <100ms, no live-scrape latency) + 通達 cross-ref to 制度 / 採択 / 行政処分. Use jpcite for pre-indexed answers + 通達 cross-ref; use tax-law-mcp for ad-hoc lookups. → /compare/tax-law-mcp/
  • vs japan-corporate-mcp (yamariki-hub/japan-corporate-mcp, 8 tools, gBizINFO + EDINET + e-Stat live API, 3 user keys required): jpcite ships pre-indexed 166,969 法人 + 13,801 適格事業者 + 1,185 行政処分 + 22,258 enforcement detail with anonymous trial (no user API key required). Use jpcite for analyst pre-screening; use japan-corporate-mcp for live regulator pulls when keys are already provisioned. → /compare/japan-corporate-mcp/

Architecture overview (4-layer)

  1. Ingest / corpus — primary-source crawl (経産省, MAFF, JFC, 総務省, NTA, e-Gov, 47 都道府県公報) → am_source with content_hash + last_verified. Aggregator domains excluded.
  2. Entity / fact graph — 504,238 canonical entities × 6,228,893 source-backed facts in autonomath.db (~16 GB SQLite, FTS5 trigram + FAISS IVF+PQ embeddings, nprobe=8 floor per PERF-23). No cross-DB ATTACH.
  3. Composition / outcome — Wave 21-94 composition tools (eligibility chain, complementary programs, simulate_application, due-diligence questions, kessan briefing, jurisdiction cross-check, application kit, industry packs construction/manufacturing/real_estate, plus Wave 60-94 agent_briefing_pack + agent_cohort_deep/ultra primitives).
  4. Wire layer — FastMCP (stdio, MCP 2025-06-18) + FastAPI REST (/v1/*). 364 stable-gate routes, 307 OpenAPI paths, 261 default-gate MCP tools. Stripe metered billing on Fly.io Tokyo + Cloudflare Pages + Cloudflare WAF. Token-bucket rate-limit middleware on every request. decision_insights / next_questions / eligibility_gaps / document_readiness / decision_support envelopes for agent-side scaffolding.

Full refresh: root AGENTS.md (vendor-neutral SOT) + DIRECTORY.md (directory map) + CLAUDE.md (Claude-specific shim).

MCP tools — 261 at default gates

Group Coverage
Core Programs, Case Studies, Loans, Enforcement, Exclusions, Laws, Court Decisions, Bids, Tax Rulesets, Quota probe (get_usage_status)
Audit / composition audit_batch_evaluate, compose_audit_workpaper, resolve_citation_chain
jpcite generic Entity/Fact DB, funding stack, evidence/source manifests, lifecycle/graph/rule-engine, tax/certification/loan/enforcement wrappers
V4 universal get_annotations, validate, get_provenance, get_provenance_for_fact
Static resources list_static_resources_am, get_static_resource_am, list_example_profiles_am, get_example_profile_am, deep_health_am
NTA corpus cite_tsutatsu, find_bunsho_kaitou, find_saiketsu, find_shitsugi
Eligibility composition apply_eligibility_chain_am, find_complementary_programs_am, program_active_periods_am, simulate_application_am, track_amendment_lineage_am
Application composition bundle_application_kit, cross_check_jurisdiction, forecast_program_renewal, match_due_diligence_questions, prepare_kessan_briefing
Industry packs pack_construction, pack_manufacturing, pack_real_estate
Corporate layer get_houjin_360_am, list_edinet_disclosures, search_invoice_by_houjin_partial
Wave 60-94 outcome / cohort agent_briefing_pack, agent_cohort_deep, agent_cohort_ultra, M&A / talent / brand / safety / real_estate / insurance outcome primitives

Default-gate tool count is canonical at 261 (scripts/distribution_manifest.yml tool_count_default_gates). Full list with arguments: docs/mcp-tools.md. Runtime probe: python scripts/probe_runtime_distribution.py.

Constraints / non-goals (what jpcite isn't)

  • Not legal / tax / 行政書士 / 司法書士 advice (弁護士法 § 72 / 税理士法 § 52 / 行政書士法 § 1の2 / 司法書士法 § 3). Responses ship _disclaimer envelopes on every 8-fence-sensitive surface (税理士法 §52・弁護士法 §72・公認会計士法 §47の2・行政書士法 §1の2・司法書士法 §3・社会保険労務士法 §27・弁理士法 §75・労働基準法 §36).
  • No LLM inside the service — no external LLM API calls in the data / evidence path. Content endpoints are generated from the corpus and deterministic application code; reasoning lives in the caller's agent. tests/test_no_llm_in_production.py enforces this gate.
  • Not real-time amendment tracking — snapshot data with partial historical diffs. Verify primary sources before any business decision.
  • No aggregator scraping — second-tier aggregator pages excluded from citation sources where detected.
  • No subscription tiers / seat fees / annual minimums — anonymous trial calls do not require signup and remain capped at 3 requests/day per IP. Zero-touch ops, solo operator.
  • Optional disabled domains — Labor-agreement (36協定) gated behind AUTONOMATH_36_KYOTEI_ENABLED (default off pending 社労士 supervision review). Healthcare and real-estate datasets disabled by default until primary-source coverage + disclaimers are ready. Experimental reasoning tools disabled by default.

Capability boundaries: docs/honest_capabilities.md.

REST API & SDKs

WARNING: The MCP package is published on PyPI; REST SDKs remain pre-release.

Python MCP package (autonomath-mcp) — package name kept for client compatibility:

pip install autonomath-mcp
# or
uvx autonomath-mcp

TypeScript / JavaScript SDK (@autonomath/sdk) — package name kept for compatibility. Public package release pending; the REST API v1 surface is the stable contract while the SDK remains pre-release. The package ships dual ESM + CJS output with .d.ts and exposes both REST (@autonomath/sdk) and MCP (@autonomath/sdk/mcp) entry points. Zero runtime dependencies (uses platform fetch).

Runnable examples

  • Python: examples/python/ — search by prefecture, check exclusions, program detail, pandas CSV export
  • TypeScript: examples/typescript/ — search, exclusions, MCP CLI, Next.js page

Self-serve dashboards & transparency

  • Dashboard (authenticated): GET /v1/me/dashboard — month-to-date spend, request count, cap state, top tools. See docs/dashboard_guide.md.
  • Amendment alerts: POST /v1/me/alerts/subscribe — subscribe by tool / law_id / program_id / industry_jsic / all, with severity gating (critical / important / info). See docs/alerts_guide.md.
  • Stats (public transparency): GET /v1/stats/coverage (per-prefecture / authority / kind program counts), GET /v1/stats/freshness (per-source source_fetched_at distribution), GET /v1/stats/usage (anonymised request volume).

Pricing — packet bank product lines

The packet bank ships 7 product lines, all metered (no seat fees, no annual minimums):

  • raw record — ¥3 per billable unit (税込 ¥3.30); normal search/detail calls are 1 unit, batch/export endpoints bill by documented fan-out units
  • Evidence Packet — ¥30 structured, source-linked, <24h-fresh bundle (ingredient the agent finishes)
  • deep / cohort bundles — up to ¥240 for the heaviest packet lines (portfolio / regulatory-impact / cohort-ultra)
  • weekly passport — flat-rate weekly access for high-frequency agents (JST week boundary)
  • First 3 requests/day free (anonymous, IP-based, JST daily reset)
  • No subscription tiers, no seat fees, no annual minimums
  • Cost preview: /v1/cost/preview for jpcite billable-unit estimates. Use Evidence Packet include_compression=true to compare caller-supplied input-context estimates with the caller baseline. Provider output/reasoning/search/cache costs remain outside jpcite.

SLA & infrastructure

  • Monthly uptime target: 99.0% on api.jpcite.com (Fly.io Tokyo + Cloudflare Pages + Cloudflare WAF). See docs/sla.md.
  • Tokushoho disclosure — full statutory disclosure under 特定商取引法 at site/tokushoho.html.
  • Spec surfaces — site/llms.txt and site/llms-full.txt (JA); site/llms.en.txt and site/llms-full.en.txt (EN) for AI-agent discovery.

Evaluation

Tool quality is publicly verifiable: see evals/ for a 79-query gold-standard suite (gold.yaml + run.py) covering 農業 / 製造 / IT / 創業 / 都道府県 / 税制 / 融資 / 採択事例 / prescreen / 行政処分 / cross-dataset / edge cases / 7 one-shot discovery tools (smb_starter_pack / deadline_calendar / subsidy_combo_finder / similar_cases / subsidy_roadmap_3yr / regulatory_prep_pack). Every expected_ids list was generated against the local evaluation snapshot; CI runs the suite on every PR. Per-tool precision table: docs/per_tool_precision.md. Run locally with .venv/bin/python evals/run.py.

Known limitations

jpcite is a public-record evidence layer, not a legal, tax, audit, credit, or filing decision service. Corpus coverage, source freshness, and field-level provenance vary by source family. Responses include source_url, source_fetched_at, known_gaps, and disclaimers where applicable so callers can verify primary sources before business decisions.

Support

License

MIT © 2026 jpcite

Launch state — 2026-05-16/17 (Wave 50 RC1 LANDED + Wave 51 + Wave 60-94)

Wave 50 RC1 = LANDED (2026-05-16). Contract layer + production deploy preflight gate substrate fully landed across 20 commits (Stream G 6 PR + cleanup PR7 + Wave 49 G2 + 73-tick revert + Wave 51 dim K-S foundational). Wave 51 tick 0 (9/9 dim K-S + L1 source-family + L2 math sweep, 11 modules, 416 tests PASS) closed in the same session. Wave 60-94 added M&A / talent / brand / safety / real_estate / insurance outcome primitives (432 cumulative outcomes). Canonical closeouts: docs/_internal/WAVE50_RC1_FINAL_CLOSEOUT_2026_05_16.md, docs/_internal/WAVE51_DIM_K_S_CLOSEOUT_2026_05_16.md, docs/_internal/AWS_CANARY_INFRA_LIVE_2026_05_16.md.

  • mypy strict: 0 errors (tick 6 71→0 achieved; new strict errors are red gate)
  • pytest: 8215+ PASS, 0 fail (collected 8628, +200+ tests landed)
  • coverage: 76%+ (tick 9)
  • production deploy readiness gate: 4/7 PASSING (CL6 audit — earlier 7/7 regressed during PERF cascade; remediation in flight)
  • preflight: 5/5 READY (Stream A 5 preflight artifacts all READY)
  • preflight_scorecard.state: AWS_CANARY_READY (operator token gate --unlock-live-aws-commands required to flip live_aws=true; scorecard runner authority only)
  • RC1 contract layer: 19 Pydantic models + 20 JSON Schema, scripts/check_schema_contract_parity.py bidirectional round-trip 0 drift
  • Release Capsule: 21 artifacts in manifest + 14 outcome contracts (¥300-¥900 band 実値 filled) + 3 inline packets
  • AWS canary infra: Phase 1-8 DONE + Phase 9 dryrun verified ($18,425 verified credit remaining post-CL16 audit; wet-run gated on operator UNLOCK)

See CHANGELOG.md for the full release walk.

Keywords: mcp, mcp-server, mcp-tools, claude, rag, agent-tools, japan, japanese, legal-tech, subsidies, grants, loans, tax, tax-incentives, corporate-registry, enforcement, evidence, citation, government, compliance, jpcite, autonomath-mcp, 補助金, 助成金, 融資, 税制優遇, 認定制度, 採択事例, 行政処分, 国税庁, e-Gov, mcp-2025-06-18

Badges

PyPI version PyPI downloads License MCP 2025-06-18 API status

Offline / mirrored copies of the same badges live in badges/ for use in environments where shields.io is unreachable.

last_updated: 2026-05-17

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured