SimplePDF Copilot: Client-Side AI Tool Calling for PDF Form Filling
SimplePDF Copilot is an AI assistant that interacts with the SimplePDF editor to fill forms, answer questions, focus fields, add fields, and delete pages. The PDF never leaves the browser — parsing, rendering, and field detection run entirely client-side. The LLM only receives the text content and your messages, sent to your chosen provider (default DeepSeek V4 Flash, rate-capped; BYOK supported for cloud or local via LM Studio).
Key details
- Tool calling executed client-side: The LLM issues tool call intents, but execution happens in the browser via iframe postMessage. This enables speed (client-to-client) and data control — you can remove the tool that exposes document content and the LLM never sees it.
- Field detection uses CommonForms by Joe Barrow with extra post-processing heuristics for accurate form field identification.
- Stack: Tanstack Start, AI SDK from Vercel, Tailwind.
- Open source: The Copilot overlay and client-side tool calling logic are open source on GitHub. SimplePDF itself (the iframe) remains proprietary.
- Privacy first: Designed for healthcare customers where PII cannot leave the browser. The PDF is never uploaded.
By default the demo sends document content to the LLM, but you can sever that connection by removing the content-exposing tool. The demo is fully functional at the link below.
📖 Read the full source: HN AI Agents
👀 See Also
Automated .xcstrings Localization with Claude Code
A new Claude Code skill automates localization of Xcode .xcstrings files with five pipeline stages: domain scanning, comment generation, translation with CLDR plurals, grammar checking, and plural fix conversion.
Head-to-head code review experiment compares three AI tools on same codebase
A video experiment tests Codex, Claude Code, and Claude Code with Sextant on identical code review tasks, with Codex verifying findings and judging which report is more valuable. The focus is on how workflow and structure affect what AI notices and prioritizes.
Skillware adds synthetic data generator with entropy scoring for local model fine-tuning
Skillware has released a new synthetic data generator skill that uses zlib compression-ratio heuristics to score output diversity, helping prevent model collapse. The tool works out-of-the-box with Ollama, supports Gemini/Anthropic for high-reasoning batches, and outputs JSON batches for .jsonl fine-tuning pipelines.
Buyer Eval: Claude skill for B2B vendor evaluation using AI agent conversations
A Claude skill that evaluates B2B software vendors by researching your company, asking domain-specific questions, and directly interrogating vendor AI agents through the Salespeak Frontdoor API. It cross-references claims against independent sources and produces evidence-based scorecards with transparent verification levels.