OpenClaw PDF reading: fix No PDF model configured

A chat model and a PDF reader are different choices

A model being able to discuss images does not establish that OpenClaw has selected it for the PDF tool. In a September 21, 2026 test on 2026.9.4, a conversational model could work with images while the PDF tool returned the configuration error above. After explicitly selecting a PDF model, the tool read both a text-based PDF and a scanned one.

The current PDF-tool reference describes a broader selection order: an explicit PDF model, an explicit image model, a suitable session or default model, then eligible providers discovered from configured authentication. It says the tool is not exposed if no usable model resolves. The older error is therefore one possible symptom, not the message every current installation will show.

Set the PDF model explicitly when selection is the problem

Choose a suitable model that your installation can already authenticate to. Use its full configured provider/model identifier, not its display name. From the Gateway’s command line, the narrow setting is:

openclaw config set agents.defaults.pdfModel.primary "provider/model"

Replace provider/model with that identifier before running the command. This changes the primary PDF model without replacing the surrounding configuration. Read the command’s result and follow any instruction it gives to apply the change. The config CLI reference documents setting an individual path; pasting a whole configuration from another installation is unnecessary.

Then ask the agent to read the same PDF again. If the error now names authentication, quota or an unavailable model, resolve that named problem. Repeating the original setting cannot repair an account that cannot use the chosen model. The model guide explains the difference between selecting a model and having access to it.

A scanned PDF needs more than text extraction

Try selecting a sentence in your PDF viewer. If you can only select a page image, the document may be a scan without a usable text layer. Extracting text alone can then return almost nothing even though the page looks readable to you.

OpenClaw’s PDF tool uses native PDF input for supported Anthropic and Google models. Other routes extract document content, including rendering low-text pages into images. A PDF can therefore be readable without the selected model accepting the original PDF bytes directly.

In the 2026.9.4 test, the text document and the scan were separate checks. Before the configuration was corrected, extracting text helped with one but did not solve the other. Afterward, the PDF tool could answer questions about both. That proves those documents were readable, not that every table or handwritten page will be extracted correctly.

Check the answer against a page you can see

Attach the document using your normal file-sharing path, then ask for something specific enough to verify: a heading on a named page, a particular row, or the relationship between two labeled figures. Open the same page yourself and compare. A fluent summary of the filename is not proof that the document was read.

Read the attached PDF. Tell me what appears on page 2, including the main heading and the first table row if there is one. Say if any part is unreadable rather than filling it in.

For a large document, begin with the relevant pages. Once the reader works, expand the request to the actual task. If the end goal is a running workbook of invoice fields, the invoice extraction guide covers that workflow rather than treating a successful PDF read as finished bookkeeping.

A file-access error is not a model error

If the tool says it cannot access a path, first establish where the uploaded file actually landed. The PDF tool has its own permitted-path rules. A path on your laptop, or an arbitrary temporary location on the server, is not necessarily an allowed input.

In the same test, a separate attempt to view a rendered page from a temporary directory was refused; the workspace copy was accessible. Changing models would not have solved that refusal. Use the attachment or permitted workspace location and retry the read without broadening filesystem permissions just to silence the error.