How to Use OpenClaw PDF Tool

The PDF tool is for document analysis without forcing the agent to pretend every PDF is just pasted text. OpenClaw can send PDFs directly to native providers when supported, or fall back to extracting text and rendering page images when needed. That gives you a practical path for reports, contracts, invoices, manuals, and research papers while keeping the tool contract explicit about inputs, limits, and failure modes.

30-second answer

Call the pdf tool with either pdf for one document or pdfs for multiple documents. Add a prompt that says what you want: summarize, compare, extract risks, identify action items, or answer questions. You can also pass pages in fallback mode, a model override, and maxBytesMb. The result comes back as text with metadata describing the resolved model, whether native mode was used, and any fallback attempts.

Where it fits

Use PDF analysis when the source document matters. It is better than copy-pasting snippets because it keeps the file reference attached to the request and lets OpenClaw choose native PDF or extraction fallback based on provider support. For high-stakes documents, ask for citations by page or section in the prompt, then verify important claims manually before acting on them.

Docs-grounded facts

• The tool supports pdf for one PDF and pdfs for multiple PDFs.
• A call can include up to 10 PDFs.
• Anthropic and Google use native PDF mode.
• Fallback mode extracts text and may render page images.
• The pages parameter is 1-based and works in fallback mode, not native mode.
• maxBytesMb defaults to configured pdfMaxBytesMb or 10.

Set it up deliberately

The tool is exposed only when OpenClaw can resolve a PDF-capable model. Resolution checks agents.defaults.pdfModel first, then imageModel, then the session/default model, while preferring auth-backed native PDF providers when available. Local paths, file URLs, HTTPS URLs, and managed inbound media references are supported. Keep workspace-only file policy in mind: local paths outside allowed roots may be rejected.

Use it safely

Do not feed private contracts or customer documents into a provider path you have not approved. Native mode sends raw PDF bytes to Anthropic or Google. Fallback mode extracts text first and may render pages to PNG images if text is sparse. In sandbox mode, remote HTTP URLs are rejected. Unsupported schemes such as ftp are rejected. These constraints are annoying in the right way: they keep document handling explicit.

Common mistakes

The easy mistake is requesting page filters with a native PDF provider. The docs say native mode does not support pages and will return a clear error. Another mistake is asking for a generic summary when what you really need is a decision: renewal risk, missing clauses, changed numbers, or action items. The prompt should name the business question, not just the file type.

Verification checklist

For any operational decision, spot-check at least three extracted claims against the original PDF. If the result says a table changed, inspect the table. If the result says a deadline exists, verify the date. Also check the tool metadata when debugging: native true means raw PDF provider mode; native false means fallback extraction. That distinction explains many page-filter and image-rendering behaviors.

Playbook angle

The OpenClaw Playbook turns PDF handling into a repeatable document intake lane: classify the file, extract decisions, route follow-ups, and record what was verified. That is where the PDF tool becomes more than a summarizer. It becomes the first step in a dependable document operations workflow.

Operator note

How to Use OpenClaw PDF Tool works best when it is written into a small runbook instead of left as tribal knowledge. Record the intended owner, the exact config surface, the channel where results should appear, the allowed inputs, the expected output, and the rollback step. OpenClaw gives agents broad tools, but the durable value comes from making each tool boring, repeatable, and auditable. I would rather have one well-scoped PDF analysis workflow that survives a restart than five clever demos nobody can safely run next week. If the runbook cannot explain when not to use it, keep refining before automation becomes default.