MCP 服务器

furlen-mcp

Turns a CSV — or a public dataset named in plain English — into an animated chart video. Every numeric claim in the generated narration is recomputed from the source rows before export, and a claim that does not reconcile blocks the export rather than shipping.

README

furlen-mcp

An MCP server for making charts. Give Claude, Cursor, or any Model Context Protocol client a table of numbers and it returns an animated chart video — and it recomputes every figure in the narration against your rows before it will export anything.

Published to the official MCP registry as io.github.amitsha86/furlen-mcp. Backed by Furlen.

Why this one is different

Most chart tooling will draw whatever number you hand it, and an AI-assisted one will happily generate a number to draw. This server won't: the story it writes is checked claim-by-claim against the source rows server-side, and a claim that doesn't reconcile blocks the export rather than shipping. You can watch that gate run at furlen.pro/proof.

That matters more through an agent than through a UI. Nobody is watching the intermediate output, so the check has to be the thing that refuses — not a human noticing.

Setup

Create a Furlen API key at furlen.pro → Settings → API keys.
Add the server to your MCP client config. Nothing to install first; npx fetches it.

{
  "mcpServers": {
    "furlen": {
      "command": "npx",
      "args": ["-y", "furlen-mcp"],
      "env": { "FURLEN_API_KEY": "sg_live_..." }
    }
  }
}

Claude Desktop: add the block above to claude_desktop_config.json and restart.
Cursor: add it to .cursor/mcp.json.
Any other MCP client: run npx -y furlen-mcp with FURLEN_API_KEY set.

Tools

`furlen_public_data`

Fetch a real public dataset by describing it — no spreadsheet needed.

"India GDP over the last 10 years" · "compare US and China population" · "Apple revenue over 6 years" · "US healthy life expectancy" · "India exports to the US"

Resolves through Furlen's adapters for the World Bank, IMF WEO, FRED, Eurostat, UN Comtrade, WHO and company filings (SEC EDGAR, ESEF, DART, EDINET). Returns the rows, a ready-to-use csv string you can pass straight to furlen_render, and provenance naming the exact indicator.

If no connected source covers the subject it says so: a 422 with code: "subject_unsupported" means the data doesn't exist here, not that the request was malformed — don't retry it reworded. Exchange rates are the canonical example.

`furlen_render`

Turns CSV text into an animated, claim-verified chart video. One call runs the whole pipeline — profile the data, find the insights, write the story, verify every claim, render — and returns a renderId.

Arg	Values	Default
`csv` (required)	Raw CSV text, header row first	—
`audience`	`investor` · `executive` · `linkedin` · `youtube` · `internal_team` · `client_report` · `student`	`linkedin`
`outputFormat`	`video_16_9` · `video_9_16` · `video_1_1`	`video_16_9`
`format`	`mp4` · `gif`	`mp4`
`resolution`	`720p` · `1080p` · `4k`	`1080p`

`furlen_render_status`

Polls a renderId. Returns status (queued → running → completed | failed), progress, and a downloadUrl when finished. Renders usually take 30–90 seconds, so poll every few seconds rather than blocking.

Example

"Here's my quarterly revenue CSV — make a LinkedIn chart video from it with Furlen."

The agent calls furlen_render, polls furlen_render_status, and hands back the download link once the video is ready.

Example: chart public data end to end

"Chart India's GDP over the last 10 years and make me a LinkedIn video."

The agent calls furlen_public_data with that request, gets back rows plus a csv string, passes the csv to furlen_render, polls furlen_render_status, and returns the download link. No file ever touches the agent's filesystem.

What it can't do yet

Worth knowing before you wire it up:

No file uploads. Data arrives as CSV text or by naming a public dataset — not as an attached spreadsheet.
Rendering is asynchronous. There is no synchronous "give me a chart back now" call.
The sources are the sources. If no connected provider publishes the figure, you get a refusal rather than an estimate. That is the design, not a gap.
It needs an API key with the data:read and renders:write scopes, and your plan's limits on resolution, watermarking and render minutes apply exactly as in the app.