Nutrient Skills

Nutrient Skills

AI agent skills for [Nutrient](https://www.nutrient.io/) APIs and SDKs. Works with Claude Code, Codex, Pi, Cursor, Gemini CLI, and [other agents](https://github.com/vercel-labs/skills#supported-agents).

Why these skills

The document extraction skills aren't wrappers around a prompt. They run a real document engine locally, and the accuracy numbers come from a public benchmark, not a set we picked ourselves:

• Fast and free. pdf-to-markdown extracts digital-born PDFs at about 0.004 seconds per page — roughly 134× docling's throughput on the public 200-document opendataloader-bench corpus, at essentially the same overall accuracy (0.89 vs 0.89; docling edges it at the third decimal). (Throughput, not a like-for-like single-document race: Nutrient converts batch-parallel while docling runs sequentially — see the method.) Free up to 1,000 documents a month, no key and no signup.
• Tuned for how agents actually work. The skills tell the agent to parse once and search with a bounded grep, reaching for ranked retrieval only when a plain grep would flood context. In our own testing that noticeably cut the token cost of answering questions over a document, at the same accuracy — the win is the instructions, not just the tool.
• Accurate when the page is hard. The licensed --vision tier adds a local machine-vision pipeline for scanned, photographed, and handwritten documents. On the same corpus it tops every accuracy metric while still running faster than docling.
• Local by default. Binaries are signed and run on macOS (Apple Silicon), Linux, and Windows. The local CLI does not upload your source PDFs to Nutrient. (Two honest caveats: extracted text you then feed to your own LLM/agent goes to that model provider like any other prompt, and the CLI downloads its binary from Nut