agentic-notebook

Turn any folder of documents into a NotebookLM-style research notebook for your AI agent — grounded answers, verifiable citations, no invented sources.

Point your agent at a folder of PDFs, books, papers, or notes. The skill indexes everything, and from then on the agent answers questions only from those sources, citing the exact file, page, and quote behind every claim — and can turn the sources into study guides, briefing docs, FAQs, timelines, mind maps, flashcards, quizzes, data tables, and podcast or video scripts.

Available Skills

This repository includes two skill variants:

Skill	Optimized For	Key Features
agentic-notebook	Frontier & large models (Claude 3.5/3.7, Gemini, GPT-4o)	High-context research, rich studio artifacts, full interactive Q&A and analysis
agentic-notebook-mini	Small local models (7–14B: Ollama, llama.cpp, LM Studio, vLLM)	Context-frugal multi-query + grep search, disk task ledger (checkpoint.py), --toc skimming, --list, size limits, strict command quotas

Both skills share the exact same .notebook/ index format — you can build an index with one and query it with the other interchangeably.

Install

Which variant? Use agentic-notebook for frontier / large models (Claude, Gemini, GPT). Use agentic-notebook-mini if you run a small local model (7–14B via Ollama, llama.cpp, LM Studio, vLLM) or otherwise work in a tight context budget. Both share the same index format, so you can switch later.

skills CLI

Works with Claude Code, Gemini CLI, Cursor, Codex, Copilot, OpenCode, and any other agent supported by the skills CLI.

npx skills add Harduex/agentic-notebook   # pick the variant when prompted

Claude Code plugin

/plugin marketplace add Harduex/agentic-notebook

/plugin install agentic-notebook@harduex        # frontier / large models
/plugin install agentic-notebook-mini@harduex   # small local models (7–14B)

Use

Ask your agent something like:

Act as a notebook over ./my-books — what do these authors disagree about on mastering loudness?

Or for local/small model setups:

Use notebook mini over ./my-notes — find all recipes and save them to a study guide.

The first run indexes the folder into a local .notebook/ workspace inside it. After that: grounded Q&A with citations you can check ([1] book.pdf · p.41 · "exact quote from the page"), honest "the sources don't cover this" answers when they don't, and one-request artifacts — "make me a study guide from these", "give me a two-host podcast script", "build a comparison table across the papers".

Mini Variant (agentic-notebook-mini) Highlights

For 7B–14B models (e.g. Ornith, Qwen, Gemma, Llama), context budget is precious. agentic-notebook-mini fixes common small-model failure modes:

• Disk Task Ledger (checkpoint.py): For exhaustive extraction tasks ("find all X"), task state lives on disk. Candidates, per-source sweep status, and labels are stored in .notebook/tasks/<slug>.json. Work can be interrupted and resumed at any point with one command.
• Fused Compact Search: Combines multiple semantic queries and exact string greps into a single deduplicated CLI call with reciprocal-rank fusion and per-source coverage indicators.
• Source Skimming (--toc): View line-by-line outlines of whole sources in cheap, paginated batches instead of loading full text into context.
• Safety Guards: Refuses full-source text dumps on large files (>2,500 words) and enforces verbatim citation checks with verify_citations.py.

What it handles

• PDF (including scanned books, via built-in OCR), EPUB, DOCX, PPTX, Markdown, HTML, CSV, plain text, code, subtitles
• Audio, video, and image sources when your agent can transcribe or read them
• Everything stays local — sources are read in place, and all outputs live in the folder's .notebook/ directory
• A .noteignore file at the folder root excludes files from indexing (simple gitignore-style globs: drafts/, *.log); adding a pattern later prunes the already-indexed matches on the next run

Repository Structure

agentic-notebook/
├── skills/
│   ├── agentic-notebook/        # Original skill for frontier / large models
│   │   ├── SKILL.md
│   │   └── scripts/
│   └── agentic-notebook-mini/   # Mini variant for small local 7-14B models
│       ├── SKILL.md
│       ├── README.md
│       └── scripts/             # Includes checkpoint.py disk task ledger
└── .claude-plugin/              # Claude Code marketplace manifests

Multilingual

Search and indexing work across scripts and languages — Cyrillic, Greek, accented Latin (accent-insensitive both ways), RTL scripts like Arabic and Hebrew, Indic scripts, and Chinese/Japanese/Korean (matched via character bigrams, chunked by character count). Non-English scanned books OCR with --ocr-lang, e.g. --ocr-lang deu+eng.

Requirements

• Python 3.9+ — the bundled scripts are standard-library first; PDF libraries are auto-detected when present
• Optional, for scanned books: tesseract (or easyocr)

The scripts are plain, readable Python and only ever write inside the target folder's .notebook/. Read them before installing — as you should with any skill that ships scripts.

License

MIT