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

· ai · Source ↗

TLDR

  • SimplePDF Copilot lets users fill, edit, and understand PDF forms by chatting with them, using client-side tool calling.

Key Takeaways

  • Client-side tool calling means document data can stay on-device when using local models; the public demo routes chat through a remote AI provider.
  • Use cases include foreign-language forms, contract review before signing, and pre-filling from CRM or EHR data via MCP/RAG.
  • SimplePDF exposes tools as WebMCP tools, enabling compatibility with Chrome’s built-in AI APIs.
  • The Copilot is positioned for organizations but the underlying architecture is a technical demo of what local-model PDF tooling can look like.

Hacker News Comment Review

  • The demo privacy framing caused confusion: the “client-side” headline implies no data leaves the device, but the public demo sends chat messages to a remote AI provider, which commenters flagged as a meaningful mismatch for PII-heavy forms.
  • Accuracy and UX were questioned: one commenter reported SSN input landing in the wrong field, and the demo blocks progress on optional fields with no way to skip or leave blank.
  • Commenters see the real value in programmatic pre-fill pipelines (CRM, EHR, MCP/RAG) rather than manual chat-to-field interaction, with some noting DIY setups using Claude and Python already cover the local-model angle.

Notable Comments

  • @nilirl: OCR+LLM pipeline for extracting data models from hundreds of PDF forms works ~90% of the time but field reliability breaks down at scale.

Original | Discuss on HN