MCP server - Scrapebento Docs

The ScrapeBento MCP server gives an agent named tools for approved public REST API operations. It is a thin wrapper: every tool calls the existing REST API with your token, so the same scopes, credits, quotas, rate limits, and errors apply. The MCP service does not access application databases directly.

Connect

The server supports only MCP protocol 2026-07-28 over stateless Streamable HTTP. Your environment’s MCP URL ends in /mcp; for local development it is:

http://localhost:8080/mcp

Configure a URL-based MCP client with this shape, adapting field names to the client:

{
  "mcpServers": {
    "scrapebento": {
      "type": "http",
      "url": "http://localhost:8080/mcp",
      "headers": {
        "Authorization": "Bearer ${SCRAPEBENTO_API_TOKEN}",
        "MCP-Protocol-Version": "2026-07-28"
      }
    }
  }
}

Create the token under API keys in the dashboard. A tool that creates work requires api.write; account and job reads require api.read. The MCP service does not accept a global SCRAPEBENTO_API_KEY configuration—each caller sends their own Bearer token. Clients send these HTTP headers:

Authorization: Bearer <customer-api-token>
MCP-Protocol-Version: 2026-07-28
Accept: application/json, text/event-stream
Content-Type: application/json

Only POST /mcp is valid. The endpoint does not use initialize, initialized notifications, stdio, sticky sessions, or Mcp-Session-Id. It normally returns JSON and can use request-scoped SSE when the protocol requires streaming.

Supported tools

Tool	What it does
`get_balance`	Read remaining credits.
`get_job`	Read an existing job by ID.
`get_usage`	Read debits and refunds for a trailing window.
`list_requests`	List recent job summaries.
`scrape_url`	Fetch a URL as markdown, HTML, JSON, metadata, screenshot, or redirect result.
`search`	Search the web.
`search_answer`	Produce a grounded answer with sources.
`research`	Run longer multi-source grounded research.
`domain_mentions`	Find a domain across Reddit, Hacker News, and news.
`seo_snapshot`	Collect SEO, PageSpeed, and technical crawl signals.
`infra_lookup`	Collect DNS, ASN, RDAP, TLS, and technographic signals.
`linkedin_employees`	Discover current employees for a company.
`hackernews_post`	Fetch a Hacker News item and comments.
`hackernews_search`	Search Hacker News stories and comments.
`fetch_rss`	Normalize an RSS or Atom feed.
`discover_sitemap`	Discover and parse sitemaps.

Tool descriptions include stability and credit cost. Experimental endpoints, service/health endpoints, and operations awaiting explicit product approval are not exposed as tools.

Results and errors

Successful calls include text for broad client compatibility and structured content for model-readable processing. If the REST API returns an error, the tool result is marked as an MCP tool error and can include:

{
  "code": "rate_limited",
  "message": "Try again later",
  "operationId": "getPublicAPIBalance",
  "httpStatus": 429,
  "requestId": "request-id",
  "retryable": true
}

Authentication failures happen at the HTTP boundary. REST stacks, credentials, and internal response fields are never copied into tool errors. Long-running operations are submitted once and polled through the public job endpoint. Cancelling the MCP call aborts its upstream REST request. If the MCP timeout expires first, the REST job can still be retrieved later with get_job.

​Connect

​Supported tools

​Results and errors

Connect

Supported tools

Results and errors