Tako MCP Server

Tako MCP Server

An MCP (Model Context Protocol) server that provides access to Tako's knowledge base and interactive data visualizations.

What is this?

This MCP server enables AI agents to:

• Search Tako's knowledge base for charts and data visualizations
• Fetch chart preview images and AI-generated insights
• Render fully interactive Tako charts via MCP-UI

Quick start

Use the hosted endpoint at https://mcp.tako.com. Three lines, no install:

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

That's it for new users. Detailed configs for Claude Code / Claude Desktop / Cursor / Windsurf are in the next section.

Looking for the Python pip install tako-mcp server, or the Docker image, or the Smithery listing? Those are the legacy Python implementation — see Self-hosting (legacy) at the bottom. They're kept for air-gapped, custom-fork, and pre-existing deployments, but the hosted Workers endpoint is where all new tool work ships first and has the current tool surface.

Hosted (Cloudflare Workers)

The fastest path: point your MCP client at https://mcp.tako.com with a Bearer token. No install, no local server.

Endpoints:

Authentication: every request needs Authorization: Bearer <TAKO_API_TOKEN>. Get a token at tako.com → account settings → API tokens.

Claude Code

export TAKO_API_TOKEN='<your-token>'

claude mcp add tako-mcp --transport http https://mcp.tako.com/mcp \
  --header "Authorization: Bearer $TAKO_API_TOKEN"

Verify with claude mcp list (should show tako-mcp connected) or /mcp inside a session.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

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

Restart Claude Desktop. Tako MCP should appear in the available tools list.

Cursor / Windsurf

Add to ~/.cursor/mcp.json (Cursor) or the equivalent Windsurf config:

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

Notes

• Tools are discovered automatically via the MCP tools/list handshake on connect — your client always sees the current tool surface, no manual list to keep in sync.
• Hosted uses Bearer auth on the connection, not the api_token per-tool-call argument shown in the self-hosted examples below. Once authenticated, tool inputs match exactly across both transports.
• Use the staging endpoint (mcp.staging.tako.com) for testing changes against an unstable build before they reach mcp.tako.com.

Consumer hosts (OAuth)

Use this if you're connecting Tako from Claude.ai or ChatGPT — the consumer chat hosts that don't accept Bearer tokens. The hosted endpoint at https://mcp.tako.com/mcp runs an OAuth 2.1 flow that signs you in with your Tako account and connects on your behalf, no JSON config or CLI required.

If you're using Claude Code, Claude Desktop, Cursor, or Windsurf, see the Bearer-auth instructions above — those clients accept a static Authorization: Bearer header and don't need OAuth.

Prerequisites

Before connecting from Claude.ai or ChatGPT:

1. Sign up or sign in at tako.com.
2. Mint an API token at tako.com → settings → API tokens.

Step 2 is mandatory: the consent flow looks up your existing token and surfaces a "Your Tako account does not have an API token yet" page if it doesn't find one. Tako does not auto-mint a token during the OAuth dance, because rotating an existing one would break any Claude Code / Cursor wiring you already have on the same account.

What you'll see during connect

The same three Tako-hosted screens appear regardless of which host (Claude.ai or ChatGPT) you're connecting from:

1. Tako sign-in page. Two options: Continue with Google or send yourself an email magic-link. Use the same identity you signed up with at tako.com.

   

2. Tako consent page. Reads "Connect [host name] to Tako — Signed in as you@example.com — Allow / Cancel". Click Allow to authorize the connection.

   

3. Bounce back to the host. The connector is now listed and tools are callable.

The host itself (Claude.ai or ChatGPT) may also display its own consent prompt before or after Tako's. That's normal — Tako confirms it's safe to share your account; the host confirms it's safe to invoke an external connector.

Claude.ai

Requires Claude.ai Pro, Max, Team, or Enterprise.

1. Open Claude.ai → Settings → Connectors.

   

2. Click Add custom connector.

   [Screenshot: Claude.ai "Add custom connector" dialog]

3. Paste https://mcp.tako.com/mcp and click Connect.

4. You'll be taken through the Tako sign-in flow described above.

5. After consent, Tako appears in your connector list as connected.

   

ChatGPT

Requires ChatGPT Pro, Business, or Enterprise. Developer Mode must be enabled.

1. Open ChatGPT → Settings → Connectors → Developer Mode and toggle it on if it isn't already.

   

2. Click Create custom connector.

   

3. Paste https://mcp.tako.com/mcp and click Connect.

4. You'll be taken through the Tako sign-in flow described above.

5. After consent, the connector is listed and ready to use.

   

Verify it's working

In a fresh conversation, ask:

Show me Tako's chart on Intel vs Nvidia headcount.

A successful response includes a chart link or an inline chart render (depending on host) within a few seconds. If you instead see an authentication error, jump to Disconnecting & re-authorizing below.

Disconnecting & re-authorizing

There are two ways to break the connection, and they have different blast radius. Pick the one that matches what you actually want.

Per-host disconnect (Claude.ai or ChatGPT settings → remove the Tako connector). Stops that host from making MCP calls. Does not revoke the underlying Tako API token. Other connected hosts — and any Claude Code / Cursor Bearer-auth wiring on the same account — keep working unchanged.

Rotate the API token at tako.com → settings → API tokens. This is the hard kill switch. Rotating creates a new token and invalidates the old one server-side, which means every previously-issued OAuth grant — across every host — stops authenticating immediately. To resume from any host, disconnect and reconnect; the new consent flow picks up your fresh token.

This kill-switch behavior is by design for v1. Per-grant scoped tokens (revoke a single host without touching the others) are tracked under TAKO-2679's known limitations.

Self-hosting (legacy Python server)

⚠️ The pip / Docker / Smithery paths described in this section are the original Python implementation in src/tako_mcp/. They are not the current canonical Tako MCP — that's the hosted Cloudflare Worker at mcp.tako.com documented above. The Python server still works and is maintained for self-hosted, air-gapped, and Smithery-marketplace deployments, but it ships an older tool surface and an older transport (SSE, per-tool api_token argument) than the hosted version. New tool work lands in workers/ first; the Python server may diverge over time.

If you don't have a specific reason to self-host, use the hosted endpoint above.

Installation

pip install tako-mcp

Or install from source:

git clone https://github.com/anthropics/tako-mcp.git
cd tako-mcp
pip install -e .

Run the Server

tako-mcp

Or with Docker:

docker build -t tako-mcp .
docker run -p 8001:8001 tako-mcp

Connect Your Agent

Point your MCP client to http://localhost:8001.

Configuration (self-hosted only)

Environment variables apply to the Python server. The hosted Worker has its own configuration baked into workers/wrangler.jsonc and is not user-tunable.

Testing the self-hosted server

python -m tests.test_client --api-token YOUR_API_TOKEN

This verifies:

• MCP handshake and initialization
• Tool discovery
• Search, images, and insights
• MCP-UI resource generation

Breaking changes (v0.2.0)

• knowledge_search → tako_search (endpoint unchanged).
• grounding → tako_answer (now backed by GA /api/v1/answer; result is {answer, cards, web_results, request_id}).
• New: tako_contents, tako_agent.

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

Available Tools

Note on the JSON examples below: these show the input shape used by the legacy Python server (api_token passed as a per-tool argument). If you're using the hosted endpoint at mcp.tako.com, drop the api_token field — auth flows via the connection-level Authorization: Bearer … header instead. Tool inputs are otherwise compatible across both transports, and your MCP client discovers the live tool surface automatically via tools/list. The hosted Worker also ships a different tool surface than the Python server: current Workers tools are tako_search, tako_answer, tako_contents, tako_agent, get_chart_image, open_chart_ui, create_chart, create_report, get_report, list_reports, export_report, and get_credit_balance.

tako_search

Search Tako's knowledge base for charts and data visualizations.

{
  "query": "Intel vs Nvidia headcount",
  "api_token": "your-api-token",
  "count": 5,
  "search_effort": "deep"
}

Returns matching charts with IDs, titles, descriptions, and URLs.

get_chart_image

Get a preview image URL for a chart.

{
  "pub_id": "chart-id",
  "api_token": "your-api-token",
  "dark_mode": true
}

get_card_insights

Get AI-generated insights for a chart.

{
  "pub_id": "chart-id",
  "api_token": "your-api-token",
  "effort": "medium"
}

Returns bullet-point insights and a natural language description.

ThinViz API - Create Custom Charts

ThinViz lets you create charts with your own data using pre-configured templates.

list_chart_schemas

List available chart templates.

{
  "api_token": "your-api-token"
}

Returns schemas like stock_card, bar_chart, grouped_bar_chart.

get_chart_schema

Get detailed info about a schema including required components.

{
  "schema_name": "bar_chart",
  "api_token": "your-api-token"
}

create_chart

Create a chart from a template with your data.

{
  "schema_name": "bar_chart",
  "api_token": "your-api-token",
  "source": "Company Reports",
  "components": [
    {
      "component_type": "header",
      "config": {
        "title": "Revenue by Region",
        "subtitle": "Q4 2024"
      }
    },
    {
      "component_type": "categorical_bar",
      "config": {
        "datasets": [{
          "label": "Revenue",
          "data": [
            {"x": "North America", "y": 120},
            {"x": "Europe", "y": 98},
            {"x": "Asia", "y": 156}
          ],
          "units": "$M"
        }],
        "title": "Revenue by Region"
      }
    }
  ]
}

Returns the new chart's card_id, embed_url, and image_url.

MCP-UI - Interactive Charts

open_chart_ui

Open an interactive chart in the UI (MCP-UI).

{
  "pub_id": "chart-id",
  "dark_mode": true,
  "width": 900,
  "height": 600
}

Returns a UIResource for rendering an interactive iframe.

Example Flow

1. User asks: "Show me a chart about Intel vs Nvidia headcount"
2. Agent calls tako_search with the query
3. Agent receives chart results with IDs
4. Agent can:
   • Call tako_answer to get a grounded prose answer
   • Call get_chart_image for a preview
   • Call open_chart_ui to render an interactive chart

Health Checks

• GET /health - Simple "ok" response
• GET /health/detailed - JSON with status and timestamp

Architecture

Tako MCP runs in two modes depending on which distribution path you chose. Both speak the same MCP tool protocol; only the transport and host differ.

Hosted mode (mcp.tako.com) — the recommended path:

AI Agent (Claude Code/Desktop, Cursor, etc.)
    ↓
  MCP Protocol (Streamable HTTP, POST /mcp)
    ↓
Cloudflare Worker  ──  Bearer auth, tool dispatch
    ↓
Tako Django API  (api.tako.com)

The Cloudflare Worker is a thin TypeScript proxy: it extracts the Bearer token, validates the MCP request, calls the appropriate Django endpoint with the user's token forwarded as X-API-Key, and returns structured tool results. Code lives in workers/ of this repo.

Self-hosted mode (pip install / Docker) — same proxy idea, run locally:

AI Agent
    ↓
  MCP Protocol (SSE)
    ↓
Local tako-mcp process  (Python, Starlette/Uvicorn)
    ↓
Tako Django API

Use this when you need to run inside a private network, modify the server, or pin a specific version. Code lives in src/tako_mcp/.

In both modes the server:

1. Authenticates with your Tako API token (Bearer header for hosted; api_token per-tool argument for SSE)
2. Translates MCP tool calls to Tako API requests
3. Returns formatted results and UI resources

MCP-UI Support

The open_chart_ui tool returns an MCP-UI resource that clients can render as an interactive iframe. The embedded chart supports:

• Zooming and panning
• Hover interactions
• Responsive resizing via postMessage
• Light and dark themes

Clients that support MCP-UI (like CopilotKit) will automatically render these resources.

MCP Registry (maintainers)

Tako is published to the official MCP Registry as a remote server under the name io.github.TakoData/tako-mcp.

• server.json (repo root) is the registry descriptor: a remote streamable-http entry pointing at https://mcp.tako.com/mcp. The registry schema does not list tools — hosts discover them at runtime via tools/list. (This is distinct from registry/server.json, the generated in-repo tool catalog used by npm run registry:gen / registry:check.)
• Publishing is automated by .github/workflows/publish-mcp.yml. It authenticates with the registry via GitHub OIDC (no secret — the io.github.TakoData/* namespace is authorized because this repo lives in the TakoData org) and runs mcp-publisher publish. The version lives in code: bump server.json's version and merge to main and it publishes automatically. A merge that touches server.json without changing the version is a no-op (the workflow skips, so the registry never sees a duplicate). Manual workflow_dispatch publishes the checked-in version on demand.
• Branded namespace (com.tako/tako-mcp) is an optional future upgrade. It requires DNS authentication: generate an Ed25519 key, add a TXT record on tako.com, and swap the workflow's login github-oidc step for login dns --domain tako.com --private-key ${{ secrets.MCP_PRIVATE_KEY }}.

License

MIT License - see LICENSE for details.

Links

• Tako - Data visualization platform
• MCP Specification - Model Context Protocol
• MCP-UI - MCP UI rendering standard