🔀 New tool interface — discover first, then execute; dataset queries no longer need SQL

Highlights

• ★🔀 Three steps instead of one list — previously every tool was listed up front and you picked; now it is discover_tools (describe what you need in plain language) → inspect_tool (see its parameters) → execute_tool (run it). Agents no longer need a long tool list in context.
• ★🔌 Please reconnect once — already-connected clients keep using the list they fetched at connect time, so they will not see the new interface. Disable and re-enable the Twinkle Hub connector in Claude Desktop (or restart it). The server does not push tool-list changes, so this step is manual.
• ★🧮 Dataset queries are structured now — no SQL. filters (eq ne gt gte lt lte in between contains), select for columns, order_by for sorting, and group_by with aggregate (count sum avg min max).
• ★⚠️ Numeric sorts need as: "number" — source values arrive as text. Measured: real-estate transactions ordered by price descending give 28.2 billion as the first row with as: "number", and 9,999,999 without it (text sorting puts "9999999" above "28287000000") — with no error raised. as: "text" is only correct for zero-padded dates such as 2026-09-09.
• ★🏛️ New parliamentary tools — legislator records (including attendance and bills proposed), the legislative history and stated rationale behind a law, and floor-vote lookup: for a given bill, who voted for, who against, and how each party voted. Sourced from the Legislative Yuan's own publications.
• ★⏳ Legacy tool names still work for now — existing integrations will not break today. They will be retired later, announced separately; please reconnect to get the new interface. Report anything broken in discussions.

Related links