edinet-util-mcp-server

An MCP server for the EDINET API v2 (金融庁), and the builder of the company-profile snapshot used by the edinet-company-profiles skill.

EDINET API v2 を使って有価証券報告書などの開示書類を扱う MCP サーバーと、上場企業のプロフィール・ 主要指標・同業他社リストをまとめたスナップショットのビルダーです。

Unofficial. This project is not affiliated with or endorsed by the Financial Services Agency (金融庁), which operates EDINET.

Status: work in progress (0.1.0.dev0). The server, the skill and the snapshot builder work; not yet published.

Layout

The server

A skill-style MCP server over stdio, with the user's own API key:

Settings (environment variables): EDINET_API_KEY (required) and EDINET_OUTPUT_DIR (default ~/Documents/EDINET), the only place the server writes. Run it with python -m edinet_util_mcp_server; Kari finds the command through the kari.mcp_servers entry point.

The skill

skill/edinet-company-profiles/ is a standard SKILL.md folder. Copy it, with its database, into ~/.agents/skills/ (Kari's main app) or import it in an Agent Desk. Its script opens the database read-only in place:

python skill/edinet-company-profiles/scripts/edinet_profiles.py compare 7203 --peers "7267,7201"

Commands: company, peers, compare, filings, sql, about (--json for structured output).

Tests

../../.venv/bin/pytest            # offline
EDINET_LIVE=1 ../../.venv/bin/pytest tests/test_live.py   # real API over stdio, key from .env
../../.venv/bin/ruff check src tests tools skill

Building the snapshot

Put the key in .env at the repository root (EDINET_API_KEY=…; the file is ignored by git), then:

python tools/build_profiles.py all

Steps, each re-runnable on its own: cohort (every listed company in the EDINET code list) → filings (every cohort filing in the daily lists of the last 18 months, or --from..--to) → profiles (key figures and listing market from each latest annual report's CSV; a company is covered when that report shows at least 3 fiscal years) → peers (top 10 per company in the same TSE industry, scored 0.5 × scale similarity + 0.5 × similarity of the report's 事業の内容 and 経営方針 text) → finish (manifest, attribution, VACUUM).

The builder needs the dev extras (pip install -e ".[dev]", which include numpy and scikit-learn for the peers step); the server itself does not.

Downloads are cached in build/edinet-cache/, so an interrupted run resumes and a rebuild makes no repeat calls. Requests go one at a time, 1 second apart, slowing down after an HTTP 429.

Data and licence

The code is MIT-licensed. Data from EDINET is used under the 公共データ利用規約（第1.0版） (PDL1.0). The snapshot is processed data, not the original filings:

• 出典：EDINET閲覧（提出）サイト（https://disclosure2.edinet-fsa.go.jp/）、PDL1.0（https://www.digital.go.jp/resources/open_data/public_data_license_v1.0）
• EDINET閲覧（提出）サイトの提出書類をもとに GotoAI が加工して作成