SimplePDF just shipped Copilot. You fill out PDF forms by chatting with them. The demo walks through an IRS W-9 form. You tell the AI what to put where, and it figures out the fields. No hunting through tiny boxes.
How it works under the hood matters. SimplePDF uses client-side tool calling. The LLM runs locally in your browser through WebLLM or Transformers.js. Libraries like pdf-lib and pdf.js handle the PDF work. The AI gets a JSON Schema with functions like fillField(fieldName, value). When you say "put my name in the right spot," it generates structured JSON that runs locally. Your data stays put.
Well, mostly. The public demo sparked debate on Hacker News. A user pointed out PII might leave your machine in the demo's setup. The author clarified this is just a demo. The real tool targets embedded, white-labeled deployments where companies control everything. The demo page itself notes that messages go to the selected AI provider. So yes, the public version sends data out.
The integration story is where this gets practical. SimplePDF Copilot connects to CRMs and EHRs through Model Context Protocol and RAG pipelines. It pulls existing data to pre-fill forms. Handles foreign-language documents too. Walks users through dense contracts.
Companies embedding this into their own products get something useful. AI-assisted form filling in their interface, with local inference keeping data private. No third-party cloud seeing your customers' tax forms.