Skip to main content
Use case

Live web access for AI agents

Give LLM agents live web access via a hosted MCP server or the /v1/chat endpoint. Structured results with confidence scores so agents know what to trust, and a source trail they can cite.

The problem with agent web access today

Most web access tools for agents return raw HTML or unstructured markdown with no signal about quality. An agent has no way to distinguish a fact from a navigation link, or a confident extraction from a guess. It treats everything as equally reliable and makes downstream errors accordingly.

The second problem is onboarding friction. If every agent deployment requires a human to sign up for an API key, agent-native products are impossible to build. You need keys that agents can provision themselves, within a rate limit, without a human in the loop.

How SuperScraper fits

One API purpose-built for agents: structured output with confidence scores, a hosted MCP server, and self-provisioning keys.

Hosted MCP server, one config block

Add the SuperScraper MCP server to your Claude Desktop or any MCP-compatible runtime with a single config block. The server exposes six tools: scrape, extract, map, search, enrich, and crawl. No SDK to install locally.

Confidence score so agents know what to trust

Every result returns _confidence (0–1) and _extraction_method. An agent can branch: quote high-confidence fields directly, flag low-confidence ones for human review, or re-scrape with a different method. Treats web data as probabilistic, not as ground truth.

Source trail for agent citations

_provenance carries the URLs that produced the result. An agent building a research report can include these directly as citations without an extra lookup step.

/v1/chat for natural-language web tasks

POST a natural-language query to /v1/chat and the intent router classifies it, selects the right data source (curated connector or ad-hoc Playwright extraction), and returns structured data with a schema preview. Good for open-ended agent tasks where the schema is not known upfront.

MCP server

Six tools (scrape, extract, map, search, enrich, crawl) for Claude Desktop and any MCP runtime

/v1/scrape

URL → clean markdown or structured data with confidence score

/v1/extract

URL + schema → structured fields with _confidence + _provenance

/v1/chat

Natural-language query → intent-routed structured result

/v1/map

Discover all URLs on a domain, for agent site navigation

/v1/keys/provision

Mint a free-tier API key programmatically, no human signup

Plug in via MCP or call the API directly

Add the MCP server in one config block, or call /v1/scrape directly with a confidence threshold your agent can act on.

claude_desktop_config.json

{
  "mcpServers": {
    "superscraper": {
      "command": "npx",
      "args": ["superscraper-mcp"],
      "env": {
        "SS_API_KEY": "your-api-key"
      }
    }
  }
}

Or call the API from your agent code and branch on confidence:

const res = await fetch('https://superscraper-production-1381.up.railway.app/v1/extract', {
  method: 'POST',
  headers: {
    'Authorization': `Bearer ${process.env.SS_API_KEY}`,
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
    url: 'https://example.com/product',
    schema: {
      productName: 'string',
      price: 'number',
      availability: 'string',
    },
  }),
});

const data = await res.json();

if (data._confidence > 0.85) {
  // High confidence, use directly
  return data;
} else {
  // Low confidence, flag for human review
  return { ...data, _needsReview: true };
}

For natural-language tasks, route through /v1/chat:

curl -X POST https://superscraper-production-1381.up.railway.app/v1/chat \
  -H "Authorization: Bearer $SS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "message": "Get me the pricing plans from stripe.com/pricing",
    "conversationId": "agent-session-abc123"
  }'

# Response includes schema preview + structured rows + cost estimate

Frequently asked questions

How do I add the SuperScraper MCP server to Claude Desktop?

Add a block to your claude_desktop_config.json under "mcpServers". Point it at the SuperScraper MCP endpoint and provide your API key. The server exposes scrape, extract, map, search, enrich, and crawl as tools Claude can call during a conversation.

Can an AI agent provision its own API key without a human signing up?

Yes. POST /v1/keys/provision returns a free-tier key with no human signup required. An agent can call this endpoint at onboarding time and store the key for subsequent calls. The free tier gives 1,000 credits per month.

What is the difference between /v1/scrape and /v1/chat for agents?

/v1/scrape fetches a specific URL and returns clean markdown or structured data. It is good when the agent knows the exact URL to fetch. /v1/chat accepts a natural-language query, the intent router classifies it, selects the right data source, and returns structured results. Use /v1/chat when the agent is working from a user instruction rather than a known URL.

How does the confidence score help agent workflows?

The _confidence field (0–1) lets an agent decide how to use a result. High confidence (>0.85): quote directly. Medium (0.5–0.85): include with a caveat. Low (<0.5): discard or flag for human review. This prevents agents from treating a low-quality extraction as ground truth and passing bad data downstream.

Give your agent the live web

Add the MCP server in one config block, or grab a free key and call the API directly. 1,000 credits a month, no card.