# 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](https://github.com/opendataloader-project/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](https://github.com/PSPDFKit/pdf-to-markdown/blob/main/docs/benchmarks.md#speed).) 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 Nutrient's CDN on first run.) Method and full tables: [pdf-to-markdown benchmarks](https://github.com/PSPDFKit/pdf-to-markdown#benchmarks) and the [live results page](https://www.nutrient.io/api/data-extraction-api/benchmarks/). ## Available Skills | Plugin | Skill | Description | |--------|-------|-------------| | [`nutrient-dws`](plugins/nutrient-dws) | `document-processor-api` | Generate, convert, assemble, OCR, redact, sign, archive, and optimize documents via the Nutrient Document Web Services API | | [`nutrient-dws`](plugins/nutrient-dws) | `document-extraction-api` | Parse documents into a structural model or Markdown via the Nutrient Data Extraction API | | [`nutrient-dws`](plugins/nutrient-dws) | `grounded-rag-ingestion` | Chunk documents into provenance-carrying JSONL (element type, page, bbox, confidence, reading order) for grounded, auditable RAG pipelines | | [`nutrient-dws`](plugins/nutrient-dws) | `dws-viewer-api` | Upload documents and mint viewer session JWTs for the Nutrient cloud Viewer API. Embeds interactive viewing, annotation, forms, and signing via cloud infrastructure — no viewing/storage servers to run (you still mint session tokens from your backend) | | [`make-pdf`](plugins/make-pdf) | `make-pdf` | Generate PDFs from Markdown or HTML — single files or whole directories — with accessible PDF/UA, archival PDF/A, and watermark outputs, plus built-in conformance verification | | [`remediate-pdf`](plugins/remediate-pdf) | `remediate-pdf` | Remediate existing PDFs: auto-tag with PDF/UA semantic structure (headings, lists, tables, reading order) via the Nutrient DWS Accessibility API (auto-tag only — verify output with the bundled verify-pdf.py) | | [`pdf-to-markdown`](plugins/pdf-to-markdown) | `pdf-to-markdown` | Extract text from PDFs as structured, semantic Markdown | | [`pdf-to-text`](plugins/pdf-to-text) | `pdf-to-text` | Extract layout-preserving plain text from PDFs | | [`query`](plugins/query) | `query` | Find the most relevant passages in an extracted document with ranked (BM-25) search | | [`nutrient-sdk-dev`](plugins/nutrient-sdk-dev) | 13 SDK skills | Build with Nutrient SDKs — Web Viewer, Document Authoring, mobile (iOS/Android/React Native/Flutter/MAUI), server (Python/Java/Node.js/.NET), self-hosted Document Engine, and AI Assistant | ## Installation ### npx skills (recommended) Install using the [Skills CLI](https://github.com/vercel-labs/skills): ```bash npx skills add pspdfkit-labs/nutrient-skills --skill document-processor-api npx skills add pspdfkit-labs/nutrient-skills --skill document-extraction-api npx skills add pspdfkit-labs/nutrient-skills --skill grounded-rag-ingestion npx skills add pspdfkit-labs/nutrient-skills --skill dws-viewer-api npx skills add pspdfkit-labs/nutrient-skills --skill make-pdf npx skills add pspdfkit-labs/nutrient-skills --skill remediate-pdf npx skills add pspdfkit-labs/nutrient-skills --skill pdf-to-markdown npx skills add pspdfkit-labs/nutrient-skills --skill pdf-to-text npx skills add pspdfkit-labs/nutrient-skills --skill query ``` The `nutrient-sdk-dev` plugin's 13 per-SDK skills install the same way (e.g. `--skill nutrient-web-sdk`). This works with Claude Code, Codex, Cursor, Gemini CLI, and [many other agents](https://github.com/vercel-labs/skills#supported-agents). To list all available skills in this repo: ```bash npx skills add pspdfkit-labs/nutrient-skills --list ``` ### Claude Code / Codex plugin marketplace Both Claude Code and Codex support the `/plugin` command: ``` /plugin marketplace add pspdfkit-labs/nutrient-skills /plugin install nutrient-dws@nutrient-skills /plugin install make-pdf@nutrient-skills /plugin install remediate-pdf@nutrient-skills /plugin install pdf-to-markdown@nutrient-skills /plugin install pdf-to-text@nutrient-skills /plugin install query@nutrient-skills /plugin install nutrient-sdk-dev@nutrient-skills ``` After installation, the plugin's skills will automatically load in all future sessions. ### Pi You can install the Nutrient skills with: ```bash pi install git:github.com/PSPDFKit-labs/nutrient-skills ``` Pi will load all skills from the packaged `plugins/*/skills` directories. If you only want to try the package without installing it, use: ```bash pi -e git:github.com/PSPDFKit-labs/nutrient-skills ``` You can still point Pi at a specific plugin's `skills/` directory or at the repo-wide `plugins/` directory in `~/.pi/agent/settings.json` or a project-local `.pi/settings.json` if you prefer manual control. ### Manual / any agent Clone the repository and point your agent at the skill directory: ```bash git clone https://github.com/pspdfkit-labs/nutrient-skills.git # Skills live under plugins//skills//SKILL.md ``` Reference `SKILL.md` directly in your agent's context, or symlink the skill directory into wherever your agent resolves skills. --- ## Repository Layout ``` .claude-plugin/ marketplace.json Marketplace catalog AGENTS.md Agent instructions (Codex, generic) CLAUDE.md Agent instructions (Claude Code) plugins/ / One directory per plugin .claude-plugin/ plugin.json Plugin manifest (Claude Code) .codex-plugin/ plugin.json Plugin manifest (Codex) skills/ / One or more skills per plugin SKILL.md Skill definition scripts/ Optional: task scripts assets/ Optional: templates, static files references/ Optional: API docs, guides ```