VISION¶
The three concepts considered¶
1. carrel — "the library desk" (CHOSEN)¶
A carrel is a private study desk in a library: your working materials, close at hand, organized your way. The product is that desk for a modern local file collection — and it treats AI agents as first-class users of the desk alongside you.
- Every command has two audiences: pretty terminal output for humans,
--jsonfor agents and scripts. - First-class context engineering:
carrel packturns files/folders into LLM-ready context (tree + contents, filters, chunking, token estimates). - A persistent desk index (SQLite FTS5) makes everything searchable, taggable, annotatable.
- The repo itself is a Claude Code plugin marketplace: install it and Claude gets slash commands, agents, and skills that drive the same CLI.
- Flagship: a Textual TUI — the desk itself: browse, inspect, act.
2. bindery — "the document craftsman's bench"¶
Bookbinding identity (quires, folios, stamps). Strong on convert/stamp/sign/proof/annotate; weaker story for search, context packing, and agents. Charming but backward-looking: it's a metaphor for making documents, not for working with them in 2026.
3. flume — "files as flows"¶
Pipeline-first: YAML recipes, watch-folders triggering multi-step transforms, a visual pipeline builder as flagship. Most ambitious flagship, highest risk (a pipeline engine + builder UI is a session on its own), and the pipeline framing buries the individually useful tools.
Scoring (uniqueness × feasibility, 1–5, against the probed environment)¶
| Concept | Uniqueness | Feasibility | Product | Notes |
|---|---|---|---|---|
| carrel | 4 | 5 | 20 | agent-native file toolkit is a fresh, current angle; every capability maps to a present tool or pure Python; TUI flagship well-scoped |
| bindery | 3 | 5 | 15 | feasible but generic "pdf toolbox" gravity; weak marketplace synergy |
| flume | 5 | 2 | 10 | pipeline engine + builder won't ship polished in one session |
Decision: carrel. It has the best synergy with the one hard requirement — a toolkit that speaks --json everywhere and packs context natively is exactly what makes Claude Code plugins genuinely useful rather than decorative. Name check 2026-07-16: PyPI carrel unclaimed (404); no dominant GitHub project (top hit ★18); no notable trademark found in a quick search.
Product principles¶
- Two audiences, one desk. Humans get rich output; agents get
--jsonand stable exit codes. Neither is an afterthought. - Wrap the masters. pandoc, poppler, qpdf, tesseract/ocrmypdf, ImageMagick, exiftool… called through one adapter layer with capability detection.
carrel doctortells you what your desk can do today and how to extend it. - The index is the memory.
carrel index+ FTS5 search, tags, and notes live in one SQLite DB per root (.carrel/), portable and inspectable. - Degrade, never crash. Missing binary → exit 3 + install hint.
- The marketplace is the product too. Plugins group by domain and delegate to the CLI — no duplicated logic.
Flagship: carrel desk (Textual TUI)¶
Three-pane desk: file tree · inspector (metadata, preview, tags/notes) · action palette (convert, ocr, pack, thumbnail… driving the same core library). Ships after the MVP CLI is green.