@stravue/mcp

@stravue/mcp

An MCP server for Stravue. It gives an AI agent a research assistant, not a screenshot API: point it at a site with an ask ("pricing and plans"), and Stravue maps the site, picks the pages that answer the ask, and returns clean full-page desktop and mobile frames with the reason each page was picked. Every capture also gets a browsable share link and a presentation-ready deck.

Runs over stdio. Works with Claude Code, Claude Desktop, Cursor, and any other MCP client.

What it can and cannot do

It captures public pages, full-page, on desktop and phone, with drafted comments and a written reason for every pick, and it exports real decks.

It will not sign in to sites, pass paywalls or human-verification checks (it reports them honestly instead), click through flows, submit forms, or reach private addresses. Captures take minutes, not milliseconds: start one with capture, then poll get_capture. Up to 30 pages per capture, 120 requests per 5 minutes per key, and captures spend the account's normal plan allowance — a capture that produces nothing refunds itself. Agent captures are tagged as machine-made and never train Stravue's page-picking judgment.

60-second setup

• Get an API key: sign in at stravue.com, open your account page, and create a key. It is shown once — copy it. Keys look like sk_stravue_live_.... Captures spend your plan's normal allowance.

• Add the server to your client.

Claude Code:

claude mcp add stravue --env STRAVUE_API_KEY=sk_stravue_live_... -- npx -y @stravue/mcp

Claude Desktop (claude_desktop_config.json) or Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "stravue": {
      "command": "npx",
      "args": ["-y", "@stravue/mcp"],
      "env": {
        "STRAVUE_API_KEY": "sk_stravue_live_..."
      }
    }
  }
}

Config is env only:

Variable	Meaning
STRAVUE_API_KEY	Required. Created on your stravue.com account page.
STRAVUE_BASE_URL	Optional. Defaults to https://stravue.com.

Tools

map_site

{ url } → the site's page URLs from its sitemap, navigation, or a browser render. Honest failure fields pass through: blocked: true means a bot wall refused the browser; redirectsTo means the site lives at another host; hadSitemap: false with few urls means the site genuinely offers no map.

capture

{ urls?, site?, ask?, project?, device?, region? } → starts a capture and returns { id, status, share_url } immediately. Two modes:

• urls (1-30): capture exactly those pages.
• site + ask: Stravue maps the site and picks the pages that answer the ask. ask is capped at 200 characters here so the same ask works across every Stravue endpoint. If nothing on the site matches the ask, the API answers 422 rather than capturing a guess.

A capture takes minutes. Poll get_capture with the returned id.

get_capture

{ id } → status, ask, share_url and the frames captured so far. Each frame carries picked_why (why the page was picked) and, when a page hit a wall, a one-word failure: challenge, blank, sparse, or http_<status>. A page that produced nothing is absent, never faked.

export_deck

{ id } → downloads the presentation-ready .pptx to a temporary file and returns the file path.

export_markdown

{ id } → the machine-oriented Markdown export, returned inline. The right way for an agent to read a capture's content.

Every tool response includes the capture's share link when one exists — it is the browsable result, made to be handed to a human.

Rates and quota

• 120 requests per key per 5 minutes → 429. Poll every 15-30 seconds.
• A capture spends your plan's normal allowance → 402 with a plain reason when it runs out. A capture that produces zero pages refunds itself.
• Errors always carry the server's own message; the server never invents a result.

License

MIT.