Reducto Document Data Extractor

Purpose

Use Reducto's hosted MCP server to turn messy documents into agent-ready structured data: parse PDFs, extract fields with citations, split large packets into constituent documents, classify document types, edit supported documents, and retrieve complete job results.

This skill intentionally uses Reducto's hosted MCP server so agents can get started without Python, uv, or a local install.

When to Use

Use this skill when the user asks to:

• Pull structured fields from invoices, contracts, insurance claims, financial filings, medical records, mortgage packets, or forms.
• Parse a PDF into text, tables, figures, bounding boxes, and layout-aware blocks.
• Split a multi-document packet into named sections.
• Classify scanned documents by type.
• Chain document operations through Reducto job IDs.
• Inspect or retrieve a previous Reducto job result.

Hosted MCP Setup

Use the hosted server at https://mcp.reducto.ai/mcp with a Reducto API key in the Authorization header.

{
  "mcpServers": {
    "reducto": {
      "type": "http",
      "url": "https://mcp.reducto.ai/mcp",
      "headers": {
        "Authorization": "Bearer $REDUCTO_API_KEY"
      }
    }
  }
}

Get API keys from https://studio.reducto.ai/api-keys.

Important hosted-server constraint: the hosted server runs in Reducto's cloud and cannot read the user's local filesystem. For upload_file, pass a public https:// URL. For local files, either pre-upload through the Reducto API and pass the returned reducto:// URL, or switch to the local MCP server.

Input URL Rules

Reducto document tool parameters accept these URL schemes:

• https:// or http://: public document URL Reducto can fetch.
• reducto://: temporary Reducto file reference returned by upload/pre-upload.
• jobid://: reference to a previous processing job, useful for chaining parse into extract, split, classify, or get-job calls.

Do not pass raw local paths, file://, or private URLs to the hosted server. If the document is private, ask the user for a public signed URL or use a local/server-side pre-upload path.

Workflow

1. Clarify the document task

Pick the narrowest operation that satisfies the user:

Prefer get_documentation before using an unfamiliar endpoint, option, schema, or result field.

2. Validate access to the document

Before starting a job, confirm:

• The document URL is public, signed, or Reducto-hosted.
• The document contains no data the user is not allowed to process through Reducto.
• The user understands Reducto usage may consume pages/credits.
• For private documents, the signed URL expiration is long enough for processing.

3. Parse once, chain by job ID

For multi-step workflows, parse once and reuse jobid://<job_id>.

parse_document("https://example.com/contract.pdf") -> job_id="abc123"
extract_data("jobid://abc123", schema={...})
split_document("jobid://abc123", ...)
get_job(job_id="abc123")

This avoids re-uploading or reparsing the same document.

4. Extract with explicit schemas

For structured extraction, ask for or draft a schema with clear field names, types, and descriptions. Include citation requirements when the user needs auditability.

For invoices, a good starting schema is:

{
  "vendor_name": "string",
  "invoice_number": "string",
  "invoice_date": "string",
  "due_date": "string",
  "currency": "string",
  "subtotal": "number",
  "tax": "number",
  "total": "number",
  "line_items": [
    {
      "description": "string",
      "quantity": "number",
      "unit_price": "number",
      "amount": "number"
    }
  ]
}

For contracts, collect the user's target fields first, such as parties, effective date, term, renewal, termination rights, payment terms, governing law, and unusual obligations.

5. Handle large, async, or truncated results

Reducto results can include:

• job_id: save this for follow-up calls.
• studio_link: use this for visual inspection/debugging.
• usage: pages and credits consumed.
• next_steps: Reducto's suggested next operation.
• result_type: "url": call get_job(job_id=...) before reading the result.
• truncated: true: call get_job(job_id=...) or rerun with a narrower page_range.

Do not summarize a truncated or URL-backed result as complete until get_job has retrieved the full materialized result.

Safety

• Never put REDUCTO_API_KEY in committed files, prompts, logs, or screenshots.
• Do not send local private files to the hosted server unless the user explicitly authorizes a public/signed upload path.
• Do not claim extracted values are authoritative unless the output includes citations or the user accepts best-effort extraction.
• Preserve document provenance: return source URL, job ID, Studio link, page ranges, and citations when available.
• If the user asks for medical, legal, financial, or compliance conclusions, return extracted evidence and suggest human review rather than making final determinations.

Demo Prompt

Use Reducto's hosted MCP server to extract vendor, invoice number, date, total, and line items from this public invoice PDF. Return JSON with citations and the Reducto Studio link. If the result is URL-backed or truncated, call get_job before answering.

Expected Output

{
  "success": true,
  "source_url": "https://example.com/invoice.pdf",
  "job_id": "abc123",
  "studio_link": "https://studio.reducto.ai/jobs/abc123",
  "usage": {
    "num_pages": 3,
    "credits": 3
  },
  "extracted_data": {
    "vendor_name": "Example Vendor",
    "invoice_number": "INV-1001",
    "total": 1234.56,
    "line_items": []
  },
  "citations": [],
  "next_steps": "Use jobid://abc123 for follow-up extraction, split, classify, or get_job."
}