Tako

Full Documentation·Get your API key·MCP Registry

Why a data-native search API?On Tako'sVerticalRTK benchmarkof real-time domain questions (finance, economics, sports, weather), Tako outperforms the next-best web search API by21%— while using~75% fewer tool calls at up to one-tenth the cost, and answering research tasks in15.5s vs 124.2sfor OpenAI web search. It reaches parity with Exa, Parallel, Nimble, and Tavily on standard web benchmarks (SimpleQA, FRAMES) and pulls ahead where structured, real-time data matters.Read the evals →

- OAuth(Claude Code plugin, Claude.ai, Claude Desktop, ChatGPT) — a browser sign-in with your Tako account; a per-host API key is minted for you automatically.
- Bearer token(config-file clients: Cursor, Windsurf, VS Code, …) —get your API keyand paste it into the header.

Claude Code installs with one command — the plugin brings the MCP connection plus Tako's bundledresearch skills:

claude plugin marketplace add TakoData/tako-mcp && claude plugin install tako@tako

gemini extensions install https://github.com/TakoData/tako-mcp

Plugin (recommended)— installs the MCP connection plus Tako's bundledresearch skillsin one step, and works immediately on the free tier — no API key to mint or manage:

claude plugin marketplace add TakoData/tako-mcp claude plugin install tako@tako

That's it — the core tools (tako_search,tako_answer,tako_available_data) work right away on the anonymous free tier. To unlock the full toolset and your own account limits, authenticate once with OAuth: run/mcpinside Claude Code, selecttako, and chooseAuthenticate. A browser opens to sign you in with your Tako account and a per-host API key is minted for you automatically (visible and revocable attako.com/console/api-keys). The same OAuth flow powers the plugin on Claude.ai — the plugin's Tako connector connects with a click, no token pasting.

export TAKO_API_TOKEN='<your-token>' claude mcp add tako-mcp --transport http https://mcp.tako.com/mcp \ --header "Authorization: Bearer $TAKO_API_TOKEN"

{ "mcpServers": { "tako-mcp": { "type": "http", "url": "https://mcp.tako.com/mcp", "headers": { "Authorization": "Bearer <your-tako-api-token>" } } } }

{ "mcpServers": { "tako-mcp": { "type": "http", "url": "https://mcp.tako.com/mcp", "headers": { "Authorization": "Bearer <your-tako-api-token>" } } } }

{ "servers": { "tako-mcp": { "type": "http", "url": "https://mcp.tako.com/mcp", "headers": { "Authorization": "Bearer <your-tako-api-token>" } } } }

gemini extensions install https://github.com/TakoData/tako-mcp

That installs the MCP connection, Tako's bundledresearch skills, and three commands:

To unlock the full toolset and your own account limits, authenticate once: run/mcp auth takoinside Gemini CLI. A browser opens to sign you in with your Tako account, and a per-host API key is minted for you automatically (visible and revocable attako.com/console/api-keys).

Manual config: if you'd rather not install the extension, or you want to pin a?tools=surfaceor use a Bearer token instead of OAuth, add to~/.gemini/settings.json:

{ "mcpServers": { "tako-mcp": { "httpUrl": "https://mcp.tako.com/mcp", "headers": { "Authorization": "Bearer <your-tako-api-token>" } } } }

{ "mcp": { "tako-mcp": { "type": "remote", "url": "https://mcp.tako.com/mcp", "enabled": true, "headers": { "Authorization": "Bearer <your-tako-api-token>" } } } }

[mcp_servers.tako-mcp] command = "npx" args = ["-y", "mcp-remote", "https://mcp.tako.com/mcp", "--header", "Authorization: Bearer <your-tako-api-token>"]

Add to Zedsettings.json(via themcp-remotebridge):

{ "context_servers": { "tako-mcp": { "source": "custom", "command": "npx", "args": ["-y", "mcp-remote", "https://mcp.tako.com/mcp", "--header", "Authorization: Bearer <your-tako-api-token>"] } } }

The consumer chat hosts don't accept Bearer tokens.claude_desktop_config.jsononly validates stdio servers, so a remote"type": "http"entry there is silently dropped — Claude Desktop connects through Connectors like Claude.ai. The hosted endpoint runs an OAuth 2.1 flow that signs you in with your Tako account and mints a per-host key for you automatically.

Claude.ai(requires Pro, Max, Team, or Enterprise)
- Open Claude.ai →Settings → Connectors
- ClickAdd custom connector
- Pastehttps://mcp.tako.com/mcpand clickConnect
- Complete the Tako sign-in flow;Takothen appears as connected

Claude Desktop(same plan requirement as Claude.ai)
- Open Claude Desktop →Settings → Connectors
- ClickAdd custom connector
- Pastehttps://mcp.tako.com/mcpand clickConnect
- Complete the Tako sign-in flow;Takothen appears as connected

ChatGPT(requires Pro, Business, or Enterprise; Developer Mode enabled)
- Open ChatGPT →Settings → Connectors → Developer Modeand toggle it on
- ClickCreate custom connector
- Pastehttps://mcp.tako.com/mcpand clickConnect
- Complete the Tako sign-in flow; the connector is then listed and ready

During connectyou'll see three Tako-hosted screens regardless of host: a sign-in page (Google, or your Tako email and password), a consent page ("Connect [host] to Tako — Allow / Cancel"), then a bounce back to the host. The host may show its own consent prompt too — that's normal.

Disconnecting.A per-host disconnect (remove the connector in host settings) stops that host only; other hosts and Bearer-auth clients keep working. To hard-kill everything, rotate your key attako.com/console/api-keys— every previously-issued grant across every host stops authenticating immediately.

On connect, the server also advertisesMCP server instructionsthat hosts like Claude.ai, Claude Desktop, and Claude Code place in the model's system prompt. They steer data and metric questions totako_searchahead of the host's built-in web search, and note thattako_searchcovers the live web too, so one call can stand in for a separate web search on mixed questions. Built-in web search remains the fallback for queries outside Tako's coverage.

# Aliases compose as a comma-separated list on the MCP URL claude mcp add tako-mcp --transport http "https://mcp.tako.com/mcp?tools=agent,visualize" \ --header "Authorization: Bearer $TAKO_API_TOKEN"

Step 1: Install or update Tako MCP If Tako MCP already exists in your config, update it to this endpoint. Run this in your terminal: claude mcp add tako-mcp --transport http "https://mcp.tako.com/mcp" --header "Authorization: Bearer $TAKO_API_TOKEN" Step 2: Add this Claude skill --- name: tako-macroeconomics description: >- Macroeconomic and demographic indicators via Tako (sources vary — FRED, BLS, OECD, BIS, IMF, World Bank, Census). Inflation (CPI/PCE), unemployment, GDP, interest and policy rates, population, and cross-country comparisons as citation-backed charts. Use for macro monitoring, economic briefings, or any "what is/was <country>'s <indicator>" question. --- # Macroeconomics (Tako) Tako serves macro indicators as interactive, citation-backed charts. All tools below live on the Tako MCP server installed in Step 1. ## What's in range - Prices: CPI (headline, core, seasonally adjusted and not), PCE price indices, country inflation rates. - Labor: unemployment (headline, U-6, harmonised), employment measures. - Output and rates: nominal and real GDP, GDP growth, policy and market interest rates. - Demographics: population, population by age band, median age (Census, World Bank). These sit naturally in this skill's scope; don't send them elsewhere. - Prediction markets: Polymarket contracts on macro outcomes are in the graph ("Polymarket odds Fed rate cut 2026" returns "How many Fed rate cuts in 2026?"). Useful when the question is genuinely about expectations, and a trap otherwise, see below. Coverage is country-keyed. Individual countries resolve well (US, China, Japan, India). Multi-country blocs are weak: "Eurozone inflation rate" returns a Polymarket contract instead of an indicator, and "Euro area CPI inflation rate" returns nothing at all. For a bloc, query member countries and aggregate yourself, or take the figure from the web results and say so. ## Pick the tool - tako_answer: the default for any "<country> <indicator>" question. Returns the value in prose with its cited chart and rows attached, so one priced call finishes the job. Relay the answer verbatim. - tako_search: reach for it when you want breadth or a chart rather than a number, e.g. scanning several countries or indicator variants to see what exists, or pulling the card when the chart is the deliverable. FRED/OECD/BIS cards do export, so search is more useful here than in the licensed-data skills, but it still costs a round trip to turn a card into a figure. - tako_available_data: FREE, and the right tool when the question is what Tako covers, or when many variants exist and you need the exact indicator name (mandatory for PCE — see below). - Cohort/ranking asks ("which G7 economy has the highest inflation right now?") → one narrow call per country in parallel, then rank. Use tako_search if you only need to see what exists, tako_answer when you need each figure. ## Query patterns (Critical) - Query is COUNTRY + INDICATOR: "US CPI inflation", "US unemployment rate", "US federal funds rate" (which resolves to the Effective Federal Funds Rate, not the FOMC target range). - Be specific — most indicators have many variants, and they carry materially different numbers. "US unemployment rate" returns BLS headline (4.2%), IMF (4.3%), OECD harmonised (4.2%), and U-6 (7.9%) together. Name the variant when intent is precise; run tako_available_data first when it isn't. - PCE is the sharpest naming trap: "US core PCE inflation" silently returns a Core CPI card. Query the year-over-year variant by name instead — "US core PCE price index % change" returns the correct "Core PCE Price Index (% Change)" card. Verify the chosen card's title actually says PCE and (% Change); a plain "PCE Price Index" card is an index level (~130 points), not a rate. Run tako_available_data FIRST here. - Parallelize multi-part asks: send each metric as its own narrow concurrent tako_search, then synthesize — not one query. - Cross-country comparison is built in: "US vs China inflation" returns a 2-series comparison card. For currency-denominated indicators (GDP, wages), a cross-country chart may plot different currencies on one axis unnormalized. State each series' currency and convert before comparing. - sources: default to BOTH ["data", "web"] (also the tool's default when omitted). The chart grounds the number, web results add the release commentary and context that make a briefing readable, and web is the built-in fallback when Tako lacks the indicator, including the bloc-level gaps above. Price is identical either way. Narrow to ["data"] only when coverage is already confirmed (tako_available_data, pinned node_ids) or in a parallel fan-out where per-call web results would swamp context. - Empty result (zero cards): do NOT reword and retry blind; every search is billed. Recover in order: (1) tako_available_data (free) for the exact indicator name; (2) if covered, ONE more search using that exact name and NO node_ids — the canonical NAME is what recovers cards; a pin is a hard filter and returned FEWER cards than the same query unpinned on 11 of 20 pairs measured, because the graph holds near-duplicate metric nodes where only one twin carries cards; (3) if not covered, stop searching Tako and answer from the web results already in the response. Never more than 2 priced searches per underlying question (in a fan-out, each country+indicator query has its own budget). Pin a METRIC node_id with strict: true only to disambiguate when an unpinned retry returned the wrong near-duplicate indicator — and if a pinned call comes back empty, drop the pin rather than concluding Tako has no data. - Empty is usually genuine non-coverage but has been observed transient, which is why the free coverage check, not a hunch, decides whether a retry is justified. Empty also means "not covered in Tako," NOT that the indicator doesn't exist; don't infer a fact from silence. ## Reading a result Every card carries a title, a description holding the headline value, and retrieval facts: whether it is exportable, its relevance, its card type, its as-of date, its nodes (the graph entities and metrics it was built from), its source name, and its chart/embed URLs. Field names depend on the response format, so the checks below name the concept and you read whichever your response carries. A markdown response prints them on a facts line (exportable · relevance · type · freshness · nodes · source · chart · embed); a JSON response uses exportable, relevance, card_type, data_freshness.data_as_of, nodes, sources[].source_name, webpage_url and embed_url. ## Pick the right card (Critical) Tako auto-renders #0, and for macro the least-specific or stalest card often ranks first. the relevance fact is unreliable: the correct card is frequently tagged Low. If the right card isn't #0, reference it by linking its title to the card's chart URL and say it is the authoritative one. Check, in order: 1. Right variant? "US CPI inflation" returns three different headline numbers at once — a BIS country card (4.2%), a FRED "Inflation Rate" series (2.9%), and FRED "CPI Inflation Rate (Seasonally Adjusted)" (3.5%). Pick the card whose title matches the variant asked for; don't average them or take whichever is first. 2. Fresh, not a stale vintage? Stale series rank high routinely: that FRED "Inflation Rate" card at #1 ends in Jan 2024, and "US federal funds rate" ranks "Fed Funds Target Rate (Historical)", a series that ends in Dec 2008, above the current one. Compare as-of dates across cards and take the freshest match. Freshness also varies by series: Core CPI runs to Jun 2026 while Core PCE stops at Jan 2026, so don't present them as the same vintage. 3. Rate, not an index level? A (% Change) card is a rate in percent; a bare "Price Index" card is index points. Confirm from the units in description. 4. An indicator, not a prediction market? A Polymarket card answers "what do traders expect," not "what was reported." "Eurozone inflation rate" returns exactly this. Only use one when the question is about expectations, and label it as market-implied odds. 5. Right country? Confirm the card's nodes names it. Country overview cards (type card) sometimes carry no nodes at all — fall back to reading the title. ## Rendering - On a comparison card, the as-of date is the QUERY date, not the data date (the US-vs-China card reported the day it was run while its latest point is May 2026). Cite the period from the card's description, not that fact. - Cite the card's source name and as-of date. The roster is wider than FRED alone: BIS, OECD, IMF, World Bank, BLS, Census and Polymarket all appear, and the same indicator arrives from different providers on different cards. Never cite from a fixed list. - Reference the chart in prose; do NOT re-post the card's image URL as a markdown image — that double-renders the inline chart. ## Examples - Single (the common case) → tako_answer {"query": "What is the current US CPI inflation rate?", "sources": ["data", "web"]} → check the cited card matches the variant asked for - Cross-country → tako_answer {"query": "How does US inflation compare with China's?", "sources": ["data", "web"]} - Chart is the deliverable → tako_search {"query": "US vs China inflation", "sources": ["data", "web"]} → embed the card - Indicator-name question (free; note the arg is q) → tako_available_data {"q": "US core PCE"} → then pin the exact name: tako_answer {"query": "US core PCE price index % change", "sources": ["data"]} - Parallel multi-metric → four calls for CPI, core CPI, core PCE (% change) and PCE (% change); tako_search with "sources": ["data"] to see what exists, tako_answer per metric when you need the values (take the "(% Change)" cards; plain "PCE Price Index" cards are index levels) - Bloc-level ask → no Eurozone card exists; query member countries in parallel and aggregate, or take the figure from the web citations and label it web-sourced ## Output (tight and structured) 1) A 1–2 line read of the indicator, referencing the intent-matched chart 2) Source name — as-of date, and say so plainly when a figure came from the web rather than a card 3) A single [Open in Tako] link built from the card's embed URL, for the card you embedded Step 3: Ask the user to restart Claude Code Ask the user to restart Claude Code so the config change takes effect.

Tako MCP is a Cloudflare Worker — a thin TypeScript proxy deployed atmcp.tako.com:

AI Agent (Claude Code/Desktop, Cursor, Claude.ai, ChatGPT, …) ↓ MCP Protocol (Streamable HTTP, POST /mcp) Cloudflare Worker ── Bearer auth / OAuth, tool dispatch ↓ X-API-Key Tako Django API (tako.com)

The Worker extracts the Bearer (or OAuth-derived) token, validates the MCP request, calls the appropriate Django endpoint with the user's token forwarded asX-API-Key, and returns structured tool results. Code lives inworkers/.

- Health check:GET /healthreturns a simpleok.

- The default tool surface istako_search,tako_answer,tako_contents,tako_available_data— plustako_visualizeon ChatGPT and Claude, the hosts that render its chart inline. Everything else is opt-in via?tools=:tako_agent(agent; ChatGPT split pairtako_agent_start/tako_agent_wait),tako_visualize(visualize; needed only on clients that don't render widgets),get_credit_balance(credits), and graph primitivestako_graph_search/tako_graph_related/tako_graph_node(graph).
- The chart-image (get_chart_image), interactive-chart (open_chart_ui), chart-creation (create_chart), and report tools (create_report,get_report,list_reports,export_report) were removed.
- The self-hosted Python server (pip install tako-mcp/ Docker) was removed in favor of the hosted Cloudflare Worker.

Update any client config or agent prompts that referenced the old tool names or the Python SSE endpoint.

Tako is published to the officialMCP Registryas a remote server underio.github.TakoData/tako-mcp.

- Full Documentation— setup, tools, and integration guides
- Evaluating a new kind of Search API— benchmarks vs. Exa, Parallel, Nimble, Tavily; why data-native search wins
- Get your API key— Tako console
- Tako— the data visualization platform
- Tako on Smithery— MCP server listing
- MCP Registry—io.github.TakoData/tako-mcp
- MCP Specification— Model Context Protocol

Research tools, including a Sqlite-backed document stash