Haunt for AI Agents: MCP, REST, and Extraction Recipes

install paths

MCP for agents. REST for apps.

Use whichever path fits the job. Same extraction service underneath.

mcp

Claude, Cursor, Windsurf, Cline

mcp config

{
  "mcpServers": {
    "haunt": {
      "command": "npx",
      "args": ["-y", "@hauntapi/mcp-server"],
      "env": { "HAUNT_API_KEY": "PASTE" }
    }
  }
}

rest

Any app or workflow

curl

curl -X POST https://hauntapi.com/v1/extract \
  -H "Content-Type: application/json" \
  -H "x-api-key: PASTE" \
  -d '{"url":"https://example.com","prompt":"Return the page title and main offer as JSON"}'

recipes

Copy the job, not the plumbing.

These are the first extraction tasks most agents need. Start here, then tighten the prompt for your target.

company

Company enrichment

Extract what a business does, who it serves, pricing signals, proof, and contact routes.

{
  "url": "https://company.example",
  "prompt": "Extract company name, offer, audience, proof points, pricing signals, and contact links as JSON."
}

pricing

Pricing page extraction

Turn plan names, limits, features, and currencies into a structured comparison.

{
  "url": "https://vendor.example/pricing",
  "prompt": "Extract plans with price, billing cadence, limits, included features, and missing caveats."
}

rag

RAG-ready page summary

Ask for headings, claims, facts, and source notes instead of dumping noisy HTML into a vector store.

{
  "url": "https://docs.example/guide",
  "prompt": "Return a clean markdown summary with headings, important facts, API names, and source caveats."
}

monitor

Competitor change check

Extract stable fields from the same page over time so your agent can compare snapshots.

{
  "url": "https://competitor.example/product",
  "prompt": "Extract product name, price, feature list, guarantee language, and visible calls to action."
}

agent rules

Good agents handle failure cleanly.

Haunt is designed to return useful status, not fake certainty. If a target requires human verification or unsupported access, your agent should branch to a human or another route.

failure-aware agent promptsafe

Use Haunt only for public or authorised pages.
If success is false, do not invent data.
If error_code is captcha_required or requires_human_verification is true,
explain that a human verification step is needed and stop the extraction branch.
Return the source URL, status, and next safe action.