Skip to main content
Glama
jgrants-util-mcp-server

jgrants-util-mcp-server

by gotoai

jgrants-util-mcp-server

An MCP server for the public API of Jグランツ (jGrants), Japan's government subsidy portal: search subsidies, read their details, and read the text of their application guidelines.

Jグランツの公開APIを使って、補助金の検索・詳細の確認・公募要領の閲覧ができる MCP サーバーです。

Unofficial. This project is not affiliated with or endorsed by the Digital Agency (デジタル庁), which operates Jグランツ. It is written from scratch against the published API specification; it is not a fork of the Digital Agency's jgrants-mcp-server.

Tools

Tool

What it does

load_guideline

The usage guideline for the whole tool set: search strategy, details, attachments, what Jグランツ covers, and the conditions of use. It takes no arguments and always returns the whole guideline.

search_subsidies

Search by one keyword plus filters (area, industry, purpose, employee count); paged.

get_subsidy_detail

One subsidy's overview, eligibility, rate, amounts, dates, portal URL and attachment list.

read_attachment

The text of a PDF attachment, in pieces, with next_offset to continue; image-only pages are listed in pages_without_text.

The server follows a skill-style layout. Its MCP instructions are one sentence saying what the tools are for and to call load_guideline first. Everything that concerns the tool set as a whole — including the conditions of use — lives in the guideline, and each tool's description covers only that tool. Agents that load many MCP servers keep only a short entry per server in their prompt, and read the guideline when they need it.

Related MCP server: Japan Data MCP

Run

Python 3.11 or later.

pip install -e .
python -m jgrants_util_mcp_server                                  # stdio, the default
jgrants-util-mcp-server --transport streamable-http --port 8000    # HTTP at /mcp

stdio is the default because a client that installed the package starts the process itself: no port, no host name, nothing to put behind a proxy. Serving it over HTTP is the deliberate case, so it is the one that has to say so.

The package declares what to run in the kari.mcp_servers entry-point group, so a client that installs the wheel can fill its own command field in — the distribution is named jgrants-util-mcp-server while the module is jgrants_util_mcp_server, and guessing one from the other is exactly the mistake that group avoids.

Behind a reverse proxy, accept the public host name as well; otherwise the server's DNS rebinding protection rejects requests whose Host is not localhost:

jgrants-util-mcp-server --transport streamable-http --host 127.0.0.1 --port 8000 \
    --allowed-host demo.mcp.gotoai.com

--host, --port and --allowed-host apply only to HTTP; given without --transport streamable-http, the server refuses to start rather than ignore them.

When bound to a non-loopback address (for example --host 0.0.0.0 on a LAN), that protection is off; restrict access with a firewall or a proxy.

In Kari

Two ways, depending on whether the server runs on the user's own machine or on a host.

As a Python package (the server runs on the user's machine). Build a wheel with python -m build --wheel, then in Kari: MCPサーバー追加 → type Python パッケージ(STDIO) → ローカル wheel ファイル → the .whl → インストール. The command field fills itself in from the entry-point group above. Kari installs it into its own bundled Python, so the user needs no Python of their own. Once it is on PyPI, the pypi.org option takes the name instead.

Over HTTP (the server runs on a host). Add a custom extension of type Streamable HTTP with the URL http://<host>:<port>/mcp, and start the server with --transport streamable-http.

Either way, give it an ASCII name such as jgrants — the name becomes part of every tool name the agent sees.

Conditions of use

These are in the guideline the agent reads; they apply to anyone presenting the results:

  • Credit the source as 「出典:Jグランツ(jGrants)」.

  • Treat the information as a reference; confirm the latest details on the subsidy's official page (portal_url) before applying.

  • Avoid excessive repeated access to the API.

The Jグランツ API's own terms are published by the Digital Agency at https://www.jgrants-portal.go.jp/open-api; they apply to whoever runs this server.

Development

pip install -e ".[dev]"
pytest                    # offline tests
JGRANTS_LIVE=1 pytest     # also the live test against the real API
ruff check src tests

License

MIT — see LICENSE.

Available Tools

4 tools
get_subsidy_detailA
Read-only

Get one subsidy's full details: overview, eligibility, subsidy rate, amounts, dates, portal_url, and a list of its attachments (read them with read_attachment).

Inputs: subsidy_id (required). Returns: object

load_guidelineA
Read-only

Usage guideline for the Jグランツ tools: how to search, read details and attachments, what Jグランツ covers, and the conditions of use. Read it before your first search.

Inputs: no named parameters. Returns: result

read_attachmentA
Read-only

Read an attachment's text (PDF, page by page) in pieces of max_chars; continue from next_offset. pages_without_text lists PDF pages that are images only. Other formats cannot be read here.

Inputs: kind (required), subsidy_id (required), index, offset, +1 more. Returns: object

search_subsidiesA
Read-only

Search Jグランツ subsidies by keyword and filters. Returns one page of summaries with each subsidy's id and portal_url; total_count and next_offset tell whether more pages exist.

Inputs: keyword (required), sort, limit, order, +6 more. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool maps to a clearly distinct step: reading usage guidance, searching subsidies, fetching full details, and reading attachments. There is no overlap in purpose, and the descriptions make the intended sequence explicit.

    Naming Consistency5/5

    All four tools use a consistent snake_case verb_noun pattern: load_guideline, search_subsidies, get_subsidy_detail, read_attachment. The naming is predictable and readable.

    Tool Count5/5

    Four tools are exactly sufficient for a read-only subsidy lookup workflow: guidance, search, detail, and attachment reading. No tool feels redundant or missing from the core scope.

    Completeness4/5

    The surface covers the main lifecycle from search to detail to attachment reading, plus onboarding guidance. Minor gaps remain around discovering available filters/categories and reading non-PDF attachments, but agents can work around these limitations.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to query Japanese public data (laws, corporations, statistics) from official government APIs, returning normalized English metadata with source attribution.
      1
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI assistants to search Japanese government subsidies and grants, including archived closed calls and recurring application periods based on historical data.
      7 npm
      1
      MIT