Tako MCP Server

START HERE for any question that wants a value, figure, or finding: ask one specific data question, get one synthesized answer grounded in the data or web tako cites. It is the only tool whose single response can finish the job: it reads the cited pages internally, inlines the cited cards' rows, and returns a coverage verdict. Retrieval hands back captions and links you must then chase, and every extra round trip re-sends the whole conversation. Best for: a single, self-contained data question with one answer. The `answer` is synthesized from the cited sources; the `cards` are its citations. Also the values channel for non-exportable cards: when a card is `exportable: false` (usually license-gated), ask here with its METRIC node id pinned and strict:true to get the figures. Reach past it only for a different job: `tako_search` for breadth recon (it locates data, it does not carry values), `tako_available_data` when the question is what Tako covers, the Answer Agent for open-ended research. Grounds over BOTH data and web by default. Run `tako_available_data` first when unsure the data exists — pass `metric` to get the entity+metric pair — then pin the METRIC node id it returns, with strict:true (an entity-only pin, or a pin without strict, does not steer retrieval). Cited cards inline their recent rows (see include_contents/preview_rows), so the series arrives with the answer; for full history or a cited page's text, call `tako_contents` on its url. Results arrive as markdown: the synthesized answer first, then its cited data cards (headline, exportable flag, node ids, a rows-count pointer) and web citations, then source notes. The cited cards' actual rows ride in structuredContent (cards[].content), not the markdown, alongside machine essentials (request_id, usage, guidance, chart-widget fields). The top cited card also renders inline as a chart on hosts that support it — do NOT re-post `image_url` or `embed_url` as a markdown image or link, or it renders twice.

Find what proprietary, continuously-updated structured data exists on something — summarized in one call. Free and fast. Ask it when the question IS coverage: what does Tako have on X, is this measure tracked at all, what is it called. Then build the real question around what comes back. NOT a required first step — for a straightforward data question, tako_search or tako_answer directly is usually right. Worth one call first when you need a measure's EXACT name: resolving a loose phrase to the canonical metric name measurably improves what the priced call retrieves (measured, 9 of 15 pairs). Works on an entity (a company, person, or place → the metrics tracked on it, e.g. Tesla) or a metric (→ the entities it is tracked across, e.g. Inflation Rate). Tips: Know the measure? Split it: q="Carnival", metric="passenger cruise days" — you get the entity+metric pair and a runnable next_call in ~0.6s. Only omit `metric` to browse everything an entity has. One metric across many entities → one metric-first call; one entity across many metrics → one entity-first call. The returned coverage list answers all of them at once — never loop one call per name. Pass `label` when you can categorize the term (company → ORG, country → GPE, person → PERSON). Each match lists the exact metric/entity names, and structuredContent.matches[].coverage.items[] pairs each name with its node id. To land on exactly one metric, pin THAT node id alone with strict:true and name the entity in the query text; the call then returns that metric's card or nothing. When the measure is known — you passed `metric`, or `q` named a metric — `next_call` is that follow-up prewritten (query + the metric node + strict) — run it verbatim. A broad entity's coverage list is capped, so it can be truncated: treat a name you don't see as UNCONFIRMED rather than absent, and fall back to the web instead of re-calling this tool to double-check. This tool confirms a name EXISTS in the graph; it cannot confirm a chart exists behind it. If next_call returns 0 cards, retry the same query WITHOUT node_ids before concluding Tako has no data — the pin is a hard filter and the data is often held under a sibling node.

Reconnaissance and chart retrieval across the live web and proprietary data: many results at once, returned as structured cards and web links, and the top card auto-renders inline as a chart. It locates data — and for `exportable: true` cards it also includes a free 20-row preview by default (`include_contents`) — but a license-gated card carries no rows at all (headline value only, via `description`), and a web result is only a snippet, not a value. For a plain "what is X", `tako_answer` is the better-suited tool: one written figure beats parsing a preview table yourself, and reaching here first for that costs an extra round trip that re-sends the whole conversation. To ask it about a card you already have, pin that card's METRIC node id ALONE (the `mt::` entry in its `nodes`) with strict:true — pinning every node id on the card, or omitting strict, does not steer retrieval. Best for: breadth — fanning out many narrow queries in parallel to see what exists across several entities or metrics; retrieving a chart card when the chart or embed is itself the deliverable; and harvesting node ids and urls to feed `tako_answer` or `tako_contents`. It is cheap and fast, and built for exactly this fan-out. Coverage spans economics, finance, company KPIs, demographics, sports, markets, weather, elections, prediction markets, website/app traffic, real estate, energy, health, and more — metrics that sound web-only (e.g. SimilarWeb-style website traffic) are in the data graph. Each query resolves one entity + one metric ("Apple revenue", "Nvidia vs AMD gross margin"); broad or compound queries ("today's sports + odds") retrieve poorly. When the question is what Tako covers, or you need a metric's exact name, run `tako_available_data` (free) instead of guessing here. Data and web come back together — treat them as one result, not an either/or. Returns: `cards` (up to `count`) with preview rows and chart URLs, plus `web_results`. To read a web result in full, call `tako_contents` on its url (web urls are always fetchable; a card's full csv needs `exportable: true`). Non-exportable cards (`exportable: false`, usually license-gated) return no rows: read the headline value from the card's `description` when it carries one, or get specific figures via `tako_answer` — pin that card's METRIC node id ALONE (the `mt::` entry in its `nodes`) with strict:true — pinning every node id on the card, or omitting strict, does not steer retrieval (each such card carries a `values_hint` saying exactly this). Results arrive as a markdown document: a Tako Data section (per card: headline, exportable flag, node ids, chart link, a rows-count pointer), then Web Results, then source notes. The cards' actual rows and the web results' snippets ride in structuredContent (cards[].content, web_results[].snippet), not the markdown, alongside machine essentials (request_id, usage, chart-widget fields).