jpcite for AI agent developers

An MCP server for Japanese subsidy / tax / loan / law data. stdio transport, no SDK, drops straight into Claude Desktop, Cursor, and ChatGPT.

89 MCP tools at default gates (protocol 2025-06-18). 14,472 programs + 2,286 case studies + 108 loans + 1,185 enforcement + 154 laws full-text + 9,484 law catalog stubs + 35 tax rulesets + 13,801 invoice registrants + an entity-fact DB (503,930 entities / 6.12M facts / 23,805 relations / 335,605 aliases). 50 req/mo free to evaluate, ¥3/req commercial.

Start in 5 minutes → All 89 tools PyPI GitHub

Why jpcite

MCP native, 89 tools

stdio + protocol 2025-06-18

FastMCP-based, stdio transport, drops into Claude Desktop and Cursor. No SDK; one uvx line to launch. Tool schemas are exposed via list_tools, and Claude picks autonomously.

REST is OpenAPI 3.1 too

FastAPI auto-generated spec

For HTTP-only environments, the REST API mirrors the same surface. The OpenAPI 3.1 spec lives at /docs/openapi/v1.json and imports into Stoplight, Postman, Insomnia, GPTs Actions, Dify, and LangChain.

50 free/mo + ¥3/req metered

Evaluate without signup, then go commercial

50 anonymous req/month per IP (JST first-of-month reset). Commercial pricing is ¥3/req (¥3.30 tax inclusive), no hard cap, no monthly fixed fee. Stripe metered billing; qualified invoices issued automatically.

Primary use cases

1. Japanese-program RAG in Claude Desktop

Add the MCP server to your config and Claude searches programs, adoptions, loans, and laws on its own. Natural-language queries like "Tokyo subsidies for construction, around 5M yen" decompose internally into search_programs + check_exclusion.

2. Build subsidy SaaS in Cursor / Copilot

Connect MCP from Cursor and the Cursor agent calls jpcite tools directly while it generates types and implementation. Importing the OpenAPI spec also gives you fully type-safe client code.

3. GPTs Actions / Dify / LangChain

Import the OpenAPI 3.1 spec as a ChatGPT GPT Action and you have a Japanese-program search GPT instantly. Dify and LangChain take it as a REST tool.

4. Compliance feature inside your own SaaS

Add "applicable subsidies", "invoice-system status", and "enforcement check" features to accounting / sales / legal SaaS. ¥3/req metered scales linearly with end-user count, with no fixed cost. White-label is fine.

5. Embed in enterprise RAG

Combine internal Confluence with jpcite in a single RAG. combined_compliance_check / trace_program_to_law chain program → statute → circular in one search.

Start in 5 minutes

curl (anonymous, 50 free/mo)

curl 'https://api.jpcite.com/v1/programs/search?\
q=DX&tier=S,A&limit=5'

# OpenAPI spec
curl https://api.jpcite.com/v1/openapi.json

No signup. JST first-of-month reset.

Python + httpx (commercial, ¥3/req)

# pip install httpx
import httpx
client = httpx.Client(
 base_url="https://api.jpcite.com/v1",
 headers={"X-API-Key": "sk_..."}
)
r = client.get("/programs/search",
 params={"q": "DX", "limit": 5})
for p in r.json["items"]:
 print(p["name"], p["source_url"])

Use httpx.AsyncClient for async.

Claude Desktop (MCP native)

// ~/Library/Application Support/
// Claude/claude_desktop_config.json
{
 "mcpServers": {
 "autonomath": {
 "command": "uvx",
 "args": ["autonomath-mcp"],
 "env": {
 "AUTONOMATH_API_KEY": "sk_..."
 }
 }
 }
}

PyPI: autonomath-mcp; uvx launches in place.

A sample of the 89 tools

Discovery (7 tools)

  • smb_starter_pack
  • subsidy_combo_finder
  • deadline_calendar
  • dd_profile_am
  • similar_cases
  • regulatory_prep_pack
  • subsidy_roadmap_3yr

Cross-dataset glue

  • combined_compliance_check
  • trace_program_to_law
  • search_tax_incentives
  • reason_answer
  • list_tax_sunset_alerts

Autonomath (33 tools)

  • search_certifications
  • list_open_programs
  • active_programs_at
  • search_acceptance_stats_am
  • search_loans_am
  • check_enforcement_am
  • get_law_article_am
  • graph_traverse
  • program_lifecycle
  • prerequisite_chain
  • query_at_snapshot

See all 89 tool schemas →

Developer pricing

¥3/req metered (¥3.30 tax inclusive). 50 anonymous req/month for evaluation, no key needed.

Tier Free quota Paid (tax excl.) Tax incl.
Anonymous (no key)50 req/mo per IP--
Authenticated (API key)-¥3 / req¥3.30 / req
Embedded in your SaaS-¥3 / req (white-label OK)¥3.30 / req

Stripe metered billing, qualified invoices issued automatically, cancel anytime. Full pricing →

FAQ (developer)

Is stdio the only MCP transport?

stdio is the primary transport (the default for Claude Desktop and Cursor). SSE / WebSocket support is planned for late 2026. The REST API has the same coverage, so use that if your environment is HTTP-only.

What's the breakdown of the 89 tools?

Programs / case-studies / loans / enforcement search and helpers; an expansion layer (laws, tax rulesets, court decisions, bids, invoice registrants, plus cross-dataset glue); the entity-fact DB; metadata tools (annotations, validation, provenance); static dataset tools (resources, profiles, deep health); and DD-support tools (lifecycle, abstract, prerequisite, graph traversal, rule engine, related programs). Two 36協定 tools are currently in preparation. Full schemas at /docs/mcp-tools/.

Does this work with Claude Desktop, Cursor, and ChatGPT?

Claude Desktop has native MCP support — drop-in. Cursor accepts MCP config. ChatGPT calls the REST API through GPTs Actions. Sample configs are in the quickstart.

What are the rate limits?

Anonymous (no key) is 50 req/month per IP, reset at JST first-of-month 00:00. Authenticated has no hard cap, UTC monthly. Spikes don't 429. Concurrent connections are effectively unlimited across MCP and HTTP.

Where's the OpenAPI spec?

/docs/openapi/v1.json publishes the OpenAPI 3.1 spec, auto-generated from FastAPI. Imports into Stoplight, Postman, Insomnia, etc.

PyPI / npm?

PyPI: autonomath-mcp (uvx-runnable). Internal import path is jpintel_mcp (legacy). No npm planned; MCP is stdio, so uvx / pipx is the standard launcher.

How does billing work when I embed in my own SaaS?

Run all requests through one of your API keys, ¥3/req metered. Reselling is allowed; there's no requirement to display the jpcite brand (white-label is fine).

Related audiences?

Same MCP/REST seen from other angles: VC / M&A advisor (DD automation), tax advisors (tax ruleset queries), administrative scriveners (permit batch queries), SMB owners (LINE-bot end users). Compare all 5 audiences.

Developer resources

Get started

PyPI: autonomath-mcp GitHub [email protected] All 89 tools

Operator: Bookyou Inc. Cancel anytime, no minimums or termination fees · Legal disclosures (JP)