OpenClaw PDF Analysis Tool: Native Document Processing at Scale

What Changed in OpenClaw 2026.3.3

Before this release, PDF processing in OpenClaw required community skills that wrapped external PDF parsing libraries. They worked, but inconsistently — some handled text extraction only, others choked on scans, and integrating AI analysis meant chaining multiple tools together.

OpenClaw 2026.3.3 changes this by making PDF analysis a first-class native tool. When you send a PDF to your agent, OpenClaw detects it automatically and routes it to the configured PDF provider — either Anthropic's Claude or Google's Gemini — depending on your setup. The AI provider handles the PDF parsing and semantic analysis in one shot, with full document context awareness.

How Native PDF Support Works

When your agent receives a PDF file:

1. File detection: OpenClaw recognizes the MIME type as application/pdf
2. Provider routing: Routes to the configured PDF analysis provider (Anthropic or Google)
3. Document parsing: The AI provider extracts text, structure, and semantic meaning from the PDF
4. AI analysis: Your agent processes the parsed content with full awareness of document structure, sections, and context
5. Response generation: The agent responds based on deep understanding of the actual document content, not just raw text extraction

This is fundamentally different from old-school PDF text extraction. The AI sees the document as a document — it understands headings, bullet points, tables (to an extent), and the relationship between sections.

Configuration

PDF analysis is configured in your OpenClaw YAML config. The relevant settings live under the pdfModel and size limits:

openclaw:
  # Primary AI model (for chat, responses, etc.)
  model: claude-sonnet-4-20250514

  # PDF analysis model (can be different from primary)
  pdfModel: claude-sonnet-4-20250514  # Anthropic Claude
  # OR
  pdfModel: gemini-2.0-flash-exp  # Google Gemini

  # PDF file size limit (in megabytes)
  pdfMaxBytesMb: 20

  # Maximum number of pages to process
  pdfMaxPages: 100

Supported PDF Models

Anthropic Claude:

• Any Claude model that supports vision (Claude 3.5 Sonnet, Claude 3 Opus, and newer)
• Best for: General-purpose document analysis, research papers, contracts, knowledge base content
• Strengths: Strong reasoning, good at handling complex documents with nuanced language

Google Gemini:

• Gemini 2.0 Flash Experimental, Gemini 1.5 Pro, and later models with document understanding
• Best for: High-volume processing, cost-sensitive workloads, documents with visual content
• Strengths: Faster processing, lower cost for large documents, strong multimodal understanding

Size Limits and Quotas

Both Anthropic and Google impose size limits on PDF uploads:

• File size: Default is 20MB. Adjust via pdfMaxBytesMb. Both providers support up to ~50MB depending on your account tier.
• Page count: Default is 100 pages. Adjust via pdfMaxPages. Processing more pages costs more and takes longer.

These limits protect you from runaway processing costs on accidentally uploaded massive files. Tune them based on your typical document sizes and your budget.

Real-World Use Cases

Research Paper Analysis

Upload an academic PDF and ask:

"Summarize the methodology, key findings, and limitations of this paper.
Focus on the statistical significance of the results."

Native PDF support lets the AI read the paper with awareness of sections like Abstract, Introduction, Methods, Results, Discussion — not as a wall of text.

Contract Review

Upload a contract PDF and ask:

"Identify any clauses related to termination, liability limitations, or
intellectual property ownership. Flag anything that seems unusual or one-sided."

The AI can parse legalistic language and cross-reference clauses across different sections of the contract.

Invoice and Receipt Extraction

Upload invoices or receipts and ask:

"Extract line items, totals, tax amounts, and vendor information into a
structured format. Flag any discrepancies or missing information."

For structured data extraction from semi-structured documents like invoices, native PDF understanding is dramatically more reliable than regex-based parsing.

Knowledge Base Augmentation

Upload product documentation, technical specs, or policy documents and ask:

"Read this product manual and create a step-by-step troubleshooting guide
for the most common issues. Include page references where relevant."

This is powerful for internal knowledge bases — upload all your docs to OpenClaw, then query across them with full semantic understanding.

Migration from Legacy PDF Skills

If you were using community PDF skills before:

1. Remove the old skill: Delete or disable the third-party PDF processing skill from your config
2. Update your config: Add pdfModel and adjust pdfMaxBytesMb / pdfMaxPages to your needs
3. Test: Send a PDF to your agent and verify it routes to the native tool instead of the old skill
4. Audit your prompts: Some skills had custom prompts tuned for their parsing quirks. You may need to adjust how you phrase PDF-related requests

Best Practices

Cost Optimization

• Use a cheaper/faster model for high-volume PDF processing (e.g., Gemini 2.0 Flash) and your primary model for reasoning-heavy tasks
• Set pdfMaxPages conservatively if most of your documents are under 50 pages — this prevents accidental processing of 500-page files
• Consider batching: Instead of processing 1000 PDFs in a loop, do them in batches with a cost checkpoint in between

Quality Optimization

• For scanned documents that are actually images, native PDF support still works — but OCR quality depends on the provider. Gemini tends to handle scanned documents better than Claude in our testing.
• If you need structured data extraction (e.g., tables to CSV), ask the AI explicitly to output in a machine-readable format (JSON, CSV, markdown table) rather than free text
• For multi-document analysis (e.g., "Compare the termination clauses across these three contracts"), upload all documents in a single message if your channel supports multiple attachments — this gives the AI full context across all files

Limitations and Gotchas

• Password-protected PDFs: Not supported. The PDF provider can't decrypt them. Remove passwords before uploading.
• Handwriting: Recognition quality varies. Claude handles some handwritten annotations, Gemini is better at full handwritten pages. Don't rely on it for critical handwritten content without verification.
• Complex tables: Both providers struggle with deeply nested or multi-header tables. For tabular data extraction, expect some manual cleanup or post-processing.
• Cost: Processing a 20-page PDF costs more than a 20-page text file. The PDF analysis is charged at your provider's document understanding rates, which are higher than pure text token rates.

Production Deployment Considerations

If you're building a service around PDF analysis (e.g., automated contract review for a law firm):

• Queue-based processing: Don't block on PDF analysis in the request loop. Accept the PDF, return a "processing" message, and analyze asynchronously. Deliver results when done.
• Retry logic: PDF APIs can time out on very large documents. Implement retries with exponential backoff.
• Fallback providers: If Anthropic is down or rate-limited, fall back to Gemini for PDF processing (or vice versa)
• Monitoring: Track PDF processing success rates, average processing time per page, and cost per document. Alert on anomalies.

What's Next

Native PDF support is a significant step toward first-class multimodal understanding in OpenClaw. Future releases will likely add support for more document types (Word, Excel, PowerPoint) and enhanced capabilities like document comparison, version diffing, and more sophisticated table extraction.

For now, the combination of Anthropic and Gemini PDF support covers the vast majority of real-world document analysis use cases — from quick invoice extraction to deep research paper synthesis.