Tako

Plugin (recomendado) — instala a conexão MCP além das habilidades de pesquisa incluídas da Tako em uma única etapa, e funciona imediatamente no nível gratuito — sem necessidade de gerar ou gerenciar chave de API:

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

Isso é tudo — o tako_search funciona imediatamente no nível gratuito anônimo. Para desbloquear o conjunto completo de ferramentas e seus próprios limites de conta, autentique-se uma vez com OAuth: execute /mcp dentro do Claude Code, selecione tako e escolha Autenticar. Um navegador abre para você fazer login com sua conta Tako e uma chave de API por host é gerada automaticamente para você (visível e revogável em tako.com/console/api-keys). O mesmo fluxo OAuth alimenta o plugin no Claude.ai — o conector Tako do plugin conecta com um clique, sem colar token.

Se você adicionou o servidor anteriormente com claude mcp add, remova-o primeiro (claude mcp remove tako-mcp) para não acabar com duas cópias de cada ferramenta.

Atualizando de uma versão anterior do plugin? Versões mais antigas pediam uma chave de API Tako na configuração do plugin; essa configuração não existe mais, então após atualizar sua conexão cai silenciosamente no nível anônimo (tako_search executa; o resto pede login) — nada gera erro, mas seus limites de conta não estão mais ativos. Execute /mcp → tako → Autenticar uma vez (ou use o botão Conectar no Claude.ai) para restaurar o acesso autenticado completo.

Ou adicione o servidor MCP diretamente (depois autentique no local via /mcp → tako → Autenticar):

claude mcp add tako --transport http https://mcp.tako.com/mcp

Verifique com claude mcp list (deve mostrar tako conectado) ou /mcp dentro de uma sessão.

Use o selo de um clique acima, ou adicione ao ~/.cursor/mcp.json — o Cursor solicita login no primeiro uso:

{
  "mcpServers": {
    "tako": {
      "type": "http",
      "url": "https://mcp.tako.com/mcp"
    }
  }
}

Adicione à configuração MCP do seu Windsurf — o Windsurf solicita login no primeiro uso:

{
  "mcpServers": {
    "tako": {
      "type": "http",
      "url": "https://mcp.tako.com/mcp"
    }
  }
}

Adicione ao .vscode/mcp.json (workspace) ou ao seu mcp.json de usuário — o VS Code solicita login no primeiro uso:

{
  "servers": {
    "tako": {
      "type": "http",
      "url": "https://mcp.tako.com/mcp"
    }
  }
}

Extensão (recomendada) instala em um único comando e funciona imediatamente no nível gratuito, sem necessidade de gerar ou gerenciar chave de API:

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

Isso instala a conexão MCP, as habilidades de pesquisa incluídas da Tako e três comandos:

Para desbloquear o conjunto completo de ferramentas e seus próprios limites de conta, autentique-se uma vez: execute /mcp auth tako dentro do Gemini CLI. Um navegador abre para você fazer login com sua conta Tako, e uma chave de API por host é gerada automaticamente para você (visível e revogável em tako.com/console/api-keys).

Configuração manual: se preferir não instalar a extensão, ou quiser fixar uma lista de permissões ?tools=, adicione ao ~/.gemini/settings.json (autentique depois com /mcp auth tako):

{
  "mcpServers": {
    "tako": {
      "httpUrl": "https://mcp.tako.com/mcp"
    }
  }
}

Para usar uma chave de API, adicione um bloco headers com "Authorization": "Bearer <key>" — mas preencha-o completamente: o Gemini substitui referências ${VAR} não definidas literalmente, e um cabeçalho Authorization malformado é rejeitado em vez de ignorado, então um token pela metade quebra a conexão onde nenhum token teria funcionado.

Adicione ao opencode.json — o OpenCode solicita login no primeiro uso:

{
  "mcp": {
    "tako": {
      "type": "remote",
      "url": "https://mcp.tako.com/mcp",
      "enabled": true
    }
  }
}

O Codex conecta-se a servidores remotos através da ponte mcp-remote, que executa o fluxo de login no seu navegador na primeira conexão:

[mcp_servers.tako]
command = "npx"
args = ["-y", "mcp-remote", "https://mcp.tako.com/mcp"]

Adicione ao settings.json do Zed (via ponte mcp-remote, que executa o fluxo de login no seu navegador na primeira conexão):

{
  "context_servers": {
    "tako": {
      "source": "custom",
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.tako.com/mcp"]
    }
  }
}

Os hosts de chat para consumidores não aceitam tokens Bearer. O claude_desktop_config.json só valida servidores stdio, então uma entrada "type": "http" remota lá é silenciosamente descartada — o Claude Desktop conecta-se através de Conectores como o Claude.ai. O endpoint hospedado executa um fluxo OAuth 2.1 que faz login com sua conta Tako e gera uma chave por host automaticamente para você.

Pré-requisitos: apenas faça login em tako.com com a identidade que você autorizará. Você não gera um token — o fluxo de consentimento cria uma chave por host (nomeada MCP: <client>, visível e revogável em tako.com/console/api-keys). Conectar um novo host nunca rotaciona a chave de outro host; a Tako remove sua chave MCP mais antiga após dez.

Claude.ai (requer Pro, Max, Team ou Enterprise)

1. Abra Claude.ai → Configurações → Conectores
2. Clique em Adicionar conector personalizado
3. Cole https://mcp.tako.com/mcp e clique em Conectar
4. Complete o fluxo de login da Tako; Tako então aparece como conectado

Claude Desktop (mesmo requisito de plano que o Claude.ai)

1. Abra Claude Desktop → Configurações → Conectores
2. Clique em Adicionar conector personalizado
3. Cole https://mcp.tako.com/mcp e clique em Conectar
4. Complete o fluxo de login da Tako; Tako então aparece como conectado

ChatGPT — instale o aplicativo Tako do diretório de aplicativos do ChatGPT (ele conecta via OAuth e usa a superfície do aplicativo em https://mcp.tako.com/mcp/chatgpt). Para adicioná-lo manualmente como conector personalizado (requer Pro, Business ou Enterprise; Modo Desenvolvedor ativado):

1. Abra ChatGPT → Configurações → Conectores → Modo Desenvolvedor e ative-o
2. Clique em Criar conector personalizado
3. Cole https://mcp.tako.com/mcp/chatgpt e clique em Conectar
4. Complete o fluxo de login da Tako; o conector então é listado e pronto

A superfície /mcp/chatgpt é somente OAuth e ajustada para o ChatGPT (widget de gráfico interativo, uma listagem fixa de cinco ferramentas; ?tools= é ignorado lá). A URL genérica /mcp também funciona lá — você recebe imagens de gráfico em vez do widget interativo.

Durante a conexão, você verá três telas hospedadas pela Tako independentemente do host: uma página de login (Google, ou seu e-mail e senha Tako), uma página de consentimento ("Conectar [host] ao Tako — Permitir / Cancelar"), e então um retorno ao host. O host pode mostrar seu próprio prompt de consentimento também — isso é normal.

Desconectando. Uma desconexão por host (remover o conector nas configurações do host) interrompe apenas aquele host; outros hosts e clientes com autenticação Bearer continuam funcionando. Para encerrar tudo de vez, rotacione sua chave em tako.com/console/api-keys — toda concessão emitida anteriormente em todos os hosts para de autenticar imediatamente.

Duas ferramentas, um passo de distância — tako_search encontra, tako_contents busca:

• Perguntas amplas ou de múltiplas partes → buscas estreitas em paralelo. Decomponha em consultas de entidade única + métrica disparadas simultaneamente — ex.: "inflação do CPI dos EUA", "inflação do CPI núcleo dos EUA", "inflação do PCE dos EUA" — e depois sintetize você mesmo.
• Não sabe o que o Tako cobre → tako_available_data primeiro. É gratuito, retorna o nome exato da métrica para buscar, e se não houver resultado ainda resta a busca web.

Pergunta específica → busca e depois busca as linhas:

1. O usuário pergunta: "Qual foi o PIB dos EUA em 2024?"
2. O agente chama tako_search; o primeiro card traz o valor de destaque e o gráfico dele
3. Para a série em si, o agente chama tako_contents na URL desse card e responde com o gráfico inline

Dados para trabalhar → tako_search em paralelo (sintetize você mesmo):

1. O usuário pergunta: "Compare a inflação do CPI, CPI núcleo, PCE e PCE núcleo dos EUA."
2. O agente dispara quatro chamadas estreitas de tako_search simultaneamente — uma por entidade + métrica
3. Cada uma retorna um card com valores de destaque (o primeiro resultado é renderizado inline como um gráfico)
4. O agente sintetiza os quatro resultados, chamando tako_contents no url de um card se precisar das linhas completas (quando o card é exportable: true)

Copie o bloco abaixo e cole no Claude Code. Ele configurará a conexão MCP e a habilidade para você.

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-financial-research
description: >-
  Use when the user asks what a company's financial metric is or was (revenue, margins, EPS, cash flow, valuation, stock price, analyst estimates), compares companies on a metric, asks who a company competes with or what data exists on it, or wants a financial chart — including private companies and crypto spot prices. Returns the figures as structured, citation-backed data from Tako (S&P Global, Fiscal.ai, Visible Alpha, Xignite and others), each with a chart. Company-level data only, not country indicators or website traffic.
---

# Financial Research (Tako)

Tako serves company financials as structured, cited data: each result is a card carrying the headline value, the underlying rows, and a chart of the series. All tools below live on the Tako MCP server installed in Step 1. The tool descriptions and every result already carry the card fields, the `sources` guidance and the zero-card recovery; this skill covers how to shape a financial query, how to check a card against the question, and how to report.

## Workflow

1. **Query as ENTITY + METRIC**: `"Nvidia revenue"`, `"Tesla free cash flow"`, with `quarterly` or `annual` to steer the period. When the user wants several metrics or companies, run one call per pair in parallel: each answer then comes back as its own card, and the checks below apply cleanly to each.
2. **Call `tako_search`** with the default sources. Web results carry the qualitative half and the facts the graph doesn't hold (earnings dates, management commentary, companies with no filings or coverage).
3. **Check the top card against the question** with the list below. The top card renders inline automatically; if a different card is the right one, link its title to its `url` and say so.
4. **Read the figure from the card's `description`.** Fetch the series with `tako_contents` on the card's `url` only when you need the rows, such as to compute a growth rate, ratio or change yourself: search retrieves reported values and derives nothing. A locked card (`exportable: false`) is a licensing wall, not an error: quote the headline and stop.
5. **Zero cards?** Follow the recovery the result states, and keep to two priced searches per question. Empty means Tako doesn't cover it, not that the fact is false: no dividend card is not "pays no dividend".
6. **Ambiguous entity?** `"Costco"` resolves to Costco Wholesale Corporation and Costco Wholesale Australia; `"Coca-Cola"` to four listed companies. If the user's intent doesn't settle it, ask before quoting a number.
7. **Discovery asks** ("who does Nvidia compete with", "what does Tako track for Tesla", "Nvidia's acquisitions")? Resolve the entity with `tako_available_data` to get its node id, then call `tako_graph_related` on it. The first call returns the relation map with counts (`rel:competes_with`, `rel:subsidiaries`, `rel:acquisitions`, `metrics`, `sources`); pass `relation` to page one and `q` to filter it, then search on the names it returns.

## Checking a card against the question

Search ranks on relevance to the words, and financial vocabulary is dense: the same query can match a level and a rate, a segment and the total, an actual and an estimate. Before quoting, confirm:

1. **Metric.** The title names the metric asked for, not an overview of several. Overview cards ("Earnings & Estimates Overview", "Ratios Overview", "Stock Overview") summarize many metrics and lead with estimate-vs-actual; use one only when the question is that broad.
2. **Scope.** The card is company-wide unless a segment or geography was asked for. Segment cards exist for every line the company reports, so they can match the same words. If only a segment card exists, say so; never pass a segment off as the total.
3. **Unit.** A rate query can match the level (operating income vs operating margin). Confirm the unit in `description`.
4. **Reported or estimate, as asked.** Analyst-estimate and consensus cards match plain metric queries, and a future `coverage_end` is how you spot one. Both are financial data: quote the estimate when the question is about forecasts or consensus and label it so; otherwise take the reported card.
5. **Entity.** `nodes` names the company asked about; related listed entities share names. Some cards (Fiscal.ai charts, Stock and Ratios Overviews) carry no nodes; fall back to the title there.

## Comparisons

- A two-series comparison card exists for many pairs but not all; some pairs return two single-entity cards. Treat a card as a comparison only if every compared entity appears in its `nodes` or title; otherwise synthesize from the per-entity cards. Comparisons default to annual; say `quarterly` for quarterly.
- A cross-currency pair plots both series on one axis unnormalized. State each currency and convert before comparing; never present the raw chart as like-for-like.
- Period labels are calendar-normalized (a September fiscal year-end shows as Dec 31). Flag the normalization when the fiscal period matters.

## Output

1. One or two lines on the finding, referencing the chart in prose. Never re-post the image URL: it double-renders the inline chart.
2. Source name and `coverage_end` date. Cite the source the card names; it varies by metric. Say plainly when a figure came from a web result rather than a card.
3. One `[Open in Tako](url)` link 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.

Copie o bloco abaixo e cole no Claude Code. Ele configurará a conexão MCP e a habilidade para você.

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-web-traffic
description: >-
  Use when the user asks how much traffic a website gets, compares sites' visits, wants a top-sites ranking, asks which sites compete with a brand for attention, or asks about an app's monthly active users — including when they name the brand rather than the domain ("how much traffic does Netflix get"). Returns SimilarWeb figures as structured, citation-backed data from Tako, each with a chart. Not for company financials such as subscribers or revenue.
---

# Web & App Traffic (Tako)

Tako serves SimilarWeb traffic data as structured, cited data: each result is a card carrying the headline value and a chart of the series. All tools below live on the Tako MCP server installed in Step 1. The tool descriptions and every result already carry the card fields, the `sources` guidance and the zero-card recovery; this skill covers how to shape a traffic query, how to check a card against the question, and how to report.

## Workflow

1. **Query by domain**: `"netflix.com monthly visits"`, `"chatgpt.com"`. Traffic data is keyed by domain, so resolve a brand to its domain yourself. For app usage, query app name + metric: `"Spotify app monthly active users"`; app cards are keyed by the bare app name.
2. **Call `tako_search`** with the default sources. Web results add competitive write-ups and ranking roundups.
3. **Check the top card against the question** with the list below. You have the traffic number only when the card's title is `<domain> Monthly Visits`, or the app's active-users card.
4. **Read the figure from `description`**: the latest monthly value and the % change over the period. Traffic cards are licensed and don't export (`exportable: false`), so the description and the chart are the data; don't call `tako_contents` on one. Web-result urls remain fetchable.
5. **Zero cards?** If the query wasn't a bare domain, make it one and retry once. If a domain query is still empty, answer from the web results and label the figure web-sourced. Don't use `tako_available_data` to rule a domain out: the graph resolves brands to companies, not to domains, so a domain it doesn't know can still have a traffic card.
6. **Competitive asks** ("who competes with Netflix for attention", "Netflix vs its rivals")? Resolve the brand with `tako_available_data` to get its company node id, call `tako_graph_related` with `relation: "rel:competes_with"`, map the competitors to their domains yourself, and run one domain search per site in parallel.

## Checking a card against the question

1. **Domain, from the title.** Traffic cards list only the metric in `nodes`; the domain never appears there, so the entity check that works elsewhere doesn't apply.
2. **Absolute vs relative.** An `"A vs B"` card's description reports each series as a % change over the period. For absolute visits, read each domain's single-series card; run a per-domain search if the comparison result didn't include them.
3. **Ranking scope.** Ranking cards exist per category and per measure ("Top Websites by Visits", "Top Arts and Entertainment Websites by Visits", "… by Average Visit Duration"). The title names both; confirm they match the question.
4. **Metric family.** SimilarWeb app "Monthly Active Users" and a company's own reported MAU are different numbers from different sources, and both can appear in one result. Say which you're quoting.
5. **Month.** `coverage_end` is the data month. Cite it.

## Output

1. One or two lines on the traffic, referencing the inline chart in prose. Never re-post the image URL: it double-renders the chart.
2. "SimilarWeb" and the `coverage_end` month. Say plainly when a figure is web-sourced.
3. One `[Open in Tako](url)` link for the top card; point at extra cards by linking their titles to their `url`.

Step 3: Ask the user to restart Claude Code

Ask the user to restart Claude Code so the config change takes effect.

Copie o bloco abaixo e cole no Claude Code. Ele configurará a conexão MCP e a habilidade para você.

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: >-
  Use when the user asks what a country's economic indicator is or was (inflation, CPI, PCE, unemployment, GDP, interest rates, population), compares countries on one, asks which indicators exist for a country or which countries belong to a bloc, or wants a macro chart or briefing. Returns the figures as structured, citation-backed data from Tako (FRED, BLS, OECD, BIS, IMF, World Bank, Census and Polymarket), each with a chart. Country-level indicators only, not company financials or website traffic.
---

# Macroeconomics (Tako)

Tako serves macro and demographic indicators as structured, cited data: each result is a card carrying the headline value, the underlying rows, and a chart of the series. All tools below live on the Tako MCP server installed in Step 1. The tool descriptions and every result already carry the card fields, the `sources` guidance and the zero-card recovery; this skill covers how to shape a macro query, how to check a card against the question, and how to report.

## Workflow

1. **Query as COUNTRY + INDICATOR**: `"US CPI inflation"`, `"Japan unemployment rate"`. Coverage is keyed by country. For a bloc or region, check coverage with `tako_available_data` before searching: member-country series are far denser, so aggregate them yourself or take the figure from the web results and say so.
2. **Name the variant when intent is precise.** Most indicators exist in several variants with materially different values: headline vs core, seasonally adjusted or not, BLS vs IMF vs OECD-harmonised, U-3 vs U-6, target rate vs effective rate. When you don't know the exact name, call `tako_available_data` first, which is free, and search on the name it returns. Level and rate series share a stem: "Core PCE Price Index" is the index level and "Core PCE Price Index (% Change)" is the inflation rate, and a query that says "inflation" can match either, or a CPI series.
3. **Call `tako_search`** with the default sources. Web results carry release commentary and cover the gaps in bloc-level data.
4. **Check the top card against the question** with the list below. If a different card is the right one, link its title to its `url` and say so.
5. **Read the value from `description`.** When the card is exportable and you need more than the headline, `tako_contents` on its `url` returns the series.
6. **Zero cards?** Follow the recovery the result states, and keep to two priced searches per question. Empty means not covered, not that the indicator doesn't exist.
7. **Discovery asks** ("what does Tako track for Japan", "which countries are in the G7", "who publishes Japan's data")? Resolve the country with `tako_available_data` to get its node id, then call `tako_graph_related` on it. The first call returns the relation map with counts (`metrics`, `part_of`, `siblings`, `sources`; `members` on a bloc node); pass `relation` to page one and `q` to filter it, then search on the names it returns.

## Checking a card against the question

One indicator name covers many series: providers, methodologies, vintages, levels and rates, and prediction markets on the same quantity. Before quoting, confirm:

1. **Variant.** The title names the variant asked for; several providers' headline numbers can come back together. Don't average them or take the first.
2. **Vintage, as asked.** For a current figure, take the series whose `coverage_end` is latest among the matches: discontinued series stay in the graph and an annual series can sit a year behind the monthly one. For a historical period, any series that covers it is valid. Series refresh on different schedules, so don't present two indicators as the same vintage without checking.
3. **Rate, not level.** A "(% Change)" card is a percentage; a bare "Price Index" card is index points. Confirm from the unit in `description`.
4. **Indicator or expectation, as asked.** Polymarket cards report what traders expect, not what a statistics agency published, and exist for many macro quantities. Quote one when the question is about expectations, and label it market-implied; otherwise take the published series.
5. **Country.** `nodes` names it. Country overview cards sometimes carry no nodes; fall back to the title there.

## Comparisons

- Cross-country comparison is built in: `"US vs China inflation"` returns a two-series card. For currency-denominated indicators (GDP, wages) the chart plots both currencies on one axis unnormalized; state each currency and convert before comparing.

## Output

1. One or two lines on the indicator, referencing the chart in prose. Never re-post the image URL: it double-renders the inline chart.
2. Source name and `coverage_end` date. The roster is wider than FRED, so cite what the card names. Say plainly when a figure is web-sourced.
3. One `[Open in Tako](url)` link 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.

v0.3.0:

• A superfície de ferramentas foi reorganizada em uma pequena listagem padrão mais adesões ?tools=, e aliases de grupo ?tools= (graph, credits, answer, visualize, agent) foram introduzidos. Tanto os aliases quanto várias dessas ferramentas foram removidos — veja a entrada acima e docs/TOOLS.md para a superfície atual.
• As ferramentas de imagem de gráfico (get_chart_image), gráfico interativo (open_chart_ui), criação de gráfico (create_chart) e relatório (create_report, get_report, list_reports, export_report) foram removidas.
• O servidor Python auto-hospedado (pip install tako-mcp / Docker) foi removido em favor do Cloudflare Worker hospedado.

Atualize qualquer configuração de cliente ou prompts de agente que referenciem os nomes antigos de ferramentas ou o endpoint SSE Python.

Tako é publicado no Registro MCP oficial como um servidor remoto sob io.github.TakoData/tako-mcp.

• server.json (raiz do repositório) é o descritor de registro: uma entrada remota streamable-http apontando para https://mcp.tako.com/mcp. O esquema não lista ferramentas — hosts as descobrem em tempo de execução via tools/list. (Distinto de registry/server.json, o catálogo de ferramentas gerado no repositório usado por npm run registry:gen / registry:check.)
• Publicação é automatizada por .github/workflows/publish-mcp.yml, autenticando via GitHub OIDC (sem segredo). A versão vive no código: aumente o version de server.json, mescle em main, e ele publica automaticamente. Uma mesclagem que não muda a versão é um no-op.
• Namespace de marca (com.tako/tako-mcp) é uma atualização futura que requer autenticação DNS (chave Ed25519 + registro TXT em tako.com).
• Versionamento & changelog são automatizados via release-please. Contribuidores usam títulos de PR Conventional Commit (squash-merge); mantenedores cortam um lançamento mesclando o PR release: X.Y.Z do bot. Veja AGENTS.md → Releases.