Show HN: Filling PDF forms with AI using client-side tool calling

Who Should Read

A frontend or backend developer who wants to add AI-assisted PDF form filling to their product.

Core Mechanics

• Chat is the interface for PDFs
  • SimplePDF Copilot is a demo for editing, filling, and understanding PDFs by chatting.
  • The demo comes preloaded with an IRS W-9 form.
• Starter prompts are built in
  • "Help me fill this form" and "What's this form about?" are offered as examples.
  • Language selection and download are available in the UI.
• The demo sends chat off-device
  • The page says chat messages go to the selected AI provider.
  • It tells users to use sample data only.
• Desktop is recommended
  • The page says it is best experienced on desktop.

Community Reactions

• The maker says it shows a local-model scenario
  • The maker says client-side tool calling with local models means no document data has to leave the user's machine.
  • Listed use cases: foreign-language forms, contract review, and pre-filling from CRM/EHR via MCP/RAG.
  • It is designed to be embedded and shipped white-labeled by customers.
• A commenter wanted clearer PII warnings
  • They said chat messages go to a remote server, so PII leaves the local machine, and this should be made clearer.
• A user reported the SSN landing in the wrong box
  • They typed an SSN and it was filled into the "4 Exemptions" box.
  • They also asked what problem this solves versus clicking and typing, and how it compares to Claude Cowork.
• A user did it locally with Claude and Python
  • Claude found the fields and wrote a Python script to insert data, with some fiddling to align things.
  • They noted the PDF never went to a remote server.
• Commenters flagged a demo flaw and accuracy worries
  • The second W-9 field (business name) is usually empty, but the demo gives no way to skip it.
  • Another commenter said OCR+LLM data-model building over hundreds of PDFs works ~90% of the time, with missed or mislabeled fields, and asked if filling errs too.

How to Apply

• Have an LLM extract a PDF's field list
  • Today, ask an LLM to output one sample PDF's field names and labels as JSON.
  • Count missed or mislabeled fields by hand and record the accuracy.
• Put empty-value rules in the tool schema
  • Mark each field of the fill tool as optional or skippable.
  • Test that it doesn't force values into fields that are usually blank.
• Show where PII is sent
  • Add text near the chat input saying which AI provider receives messages.
  • List a design that routes sensitive data through local models or client-side tool calling.
• Add a confirmation step after filling
  • After the tool fills values, show each field's content and require user approval.
  • Block submission of sensitive fields like SSN until approved.

Try It

# Paste into Claude or ChatGPT (a simplified take on field extraction/filling, not SimplePDF's own method)

Below is the text of a PDF form. Find all input fields and output them as a JSON array.
Each item: {"field": field name, "label": on-screen label, "required": true/false, "skippable": whether it can be left blank}
Then propose values for each field from the user info below. If you are unsure which box fits, do not fill it; mark it "unsure".
Do not force values into fields that can be blank.

[Form text]
{{document}}

[User info (sample data only)]
{{user_info}}

Terminology

Related Papers

• GPT-5.6 Sol can beat Factorio on peaceful mode

  GPT-5.6 Sol in Codex at low thinking finished a peaceful Factorio run with a rocket launch

• AI isn't replacing SREs, it's deskilling them

  Letting AI handle 95% of incidents may dull the skills SREs need for the critical 5%.

• Backdoor in Claude Code for VS Code extension on the Cursor Marketplace

  A tampered Claude Code extension on Cursor Marketplace planted an SSH backdoor

• Building an AI-native engineering org

  Automating each shipping step raised merges per engineer about 5x with reverts under 0.5%

• We ran AI agent swarms in Minecraft and Mindustry

  We put 50 agents in Minecraft and they invented ration ledgers and bureaucracy.

• The Agent SDK

  Cursor released Agent SDK, letting you define agents in code and deploy them with one command.

• Here be monsters: What do you still need to know when agents do the work?

  Owning agent-written code depends on fundamentals, breadth, and diving deep just in time.

• AI promised less work. I've never worked more

  Running agents in parallel brought my GitHub activity back to full-time engineer levels, and I'm exhausted.

Related Resources

• https://copilot.simplepdf.com/?share=a7d00ad073c75a75d493228e6ff7b11eb3f2d945b6175913e87898ec96ca8076&form=w9&lang=en
• https://developer.chrome.com/docs/ai/built-in