Kokkai Diet MCP server

@pipeworx/kokkai-diet

Speech-level full-text search of Japanese National Diet (parliament) proceedings back to 1947 — both chambers, every recorded speech with speaker, party, session, date, and a permalink.

Part of Pipeworx — an MCP gateway connecting AI agents to 1576+ live data sources.

Tools

• searchspeeches(query?, speaker?, speakergroup?, house?, from?, until?, startrecord?, maxrecords?) — full-text speech search by keyword and/or speaker, with optional party/house/date-range filters. Paginated (max 100/request).
• searchmeetings(query?, meetingname?, house?, from?, until?, startrecord?, maxrecords?) — find meetings/sittings by name, chamber, or date range. Paginated (max 10/request — lower cap than speeches).
• getspeech(speechid) — fetch one speech's full text and metadata by its speechID.

Auth

Keyless.

Data sources

• — speech-level full-text search.
• — meeting-level search.

Both are JSON/XML endpoints of the National Diet Library's 国会会議録検索システム (Kokkai Kaigiroku Search System).

Traps for the next person

• The API requires a real search anchor. from/until/nameOfHouse alone are accepted (200 OK) but silently return zero records — they're filters, not search conditions. searchspeeches requires query and/or speaker; searchmeetings requires query, meeting_name, or house. A no-condition request 400s with (19007)検索条件を指定してください.
• Per-request record caps differ by endpoint, and out-of-range values 400 rather than clamp. /api/speech: maximumRecords 1–100 (default 30). /api/meeting: maximumRecords 1–10 (default 3) — easy to miss since it's a third the speech cap. Both tools here clamp client-side to the valid range, but a caller reading raw upstream docs will hit (19005)/(19006) if they don't.
• Never let a capped page read as the complete result. Both tools return nextstartrecord (from the API's nextRecordPosition) and iscompleteresult (true only when nextstartrecord is null). total (upstream numberOfRecords) is the true count — a caller must check it against returned before assuming they've seen everything.
• Queries are Japanese-first. The full-text index is Japanese; partial-match works on Japanese substrings. English keywords (e.g. "semiconductors") will under-match or return zero even when substantial Japanese-language debate exists on the topic — use Japanese keywords (e.g. "半導体"). Tool descriptions say this explicitly so an empty result isn't misread as "no debate on this".
• searchspeeches returns a speech preview (400 chars), not the full text — fetch the complete speech with getspeech(speechid) using the speechid from the search hit. This keeps list responses from blowing up on speeches that run thousands of characters.
• 400 error bodies from the upstream carry a useful Japanese message + details array (e.g. (19005)maximumRecordsには1～100の値を指定してください) — this pack surfaces that text verbatim rather than a generic HTTP error, since it tells the caller exactly which parameter is wrong.

Quick Start

Add to your MCP client (Claude Desktop, Cursor, Windsurf, etc.):

{
  "mcpServers": {
    "kokkai-diet": {
      "url": "https://gateway.pipeworx.io/kokkai-diet/mcp"
    }
  }
}

What this endpoint actually serves

tools/list at https://gateway.pipeworx.io/kokkai-diet/mcp returns the tools in the table above plus the shared Pipeworx meta-tools — askpipeworx, discovertools, search_within, remember/recall and the rest of the gateway-wide set. So the tool count you see is larger than this table: a single-pack endpoint currently lists roughly 30 shared tools alongside the pack's own. The connection's initialize response states its exact scope, and is the authoritative answer for a given day.

Shortened. The full README is on GitHub.