Jak używać
- Skopiuj polecenie instalacji i sprawdź źródła przed instalacją we własnym środowisku agenta.
- W oryginale sprawdź wymagane narzędzia, dane dostępowe i obsługiwane środowiska.
- Opisz dane wejściowe i oczekiwany rezultat; porównaj wynik ze źródłami.
npx skills add https://github.com/mroxso/artcraft-skills --skill pdfcraftPrzed rozpoczęciem
- Instalacja nie obejmuje subskrypcji modelu, limitu API ani uprawnień do plików.
- Gwiazdki i instalacje pokazują aktywność, a nie niezależny test jakości lub bezpieczeństwa.
Treść z repozytorium z zachowaniem nagłówków, przykładów, kodu, tabel, linków i obrazów.
PdfCraft
Operate PdfCraft through its structured automation interfaces. Preserve the user's requested document and editing workflow.
References
PdfCraft documents automation in two bundled snapshots, captured from upstream on 2026-11-09:
- references/mcp.md — the MCP server: startup flags, tool surface, and file-access confinement. Source: PdfCraft MCP conventions.
- references/cli-batch.md —
pdfcraft-cli batchfor convert, OCR, and optimize over files and folders. Source: PdfCraft batch docs.
Both are distributed under the bundled MIT license. Relative links inside the snapshots resolve against https://github.com/storytold/pdfcraft/blob/main/docs/. The desktop control channel is described in the upstream README; discover installed commands and flags rather than borrowing schemas from other ArtCraft applications.
Select the surface
- MCP (headless, opt-in):
pdfcraft-cli mcp [--root DIR] [--compact] [--automation-read-root DIR] [--automation-write-root DIR]runs over stdin/stdout JSON lines, opens no port, and stops on EOF. It always runs in its own process; it never attaches to the desktop app. - Desktop control (live window): start
pdfcraft --control ~/.pdfcraft-control.jsonto let an agent see and operate the real interface — widget tree, clicks, typing, keys, commands, screenshots. Off by default, loopback only, and every connection must present the random token in that file. Keep the control file in a folder only you can write (never/tmp);pdfcraft-cli uirefuses symlinked or world-readable control files. Drive it withpdfcraft-cli ui --control <file> inspect|click|key|command|drag|screenshot|resize|focus|quit. - One-off file processing:
pdfcraft-cli batch(see the batch reference) and the CLI's other subcommands.
Discover available executables and sessions first and reuse a running session when possible. If the required connection or executable is missing, report the concrete prerequisite; this skill does not install PdfCraft or register MCP servers.
Tools and discovery
The MCP surface maps document work to five common tools: command_list {doc?, filter?, enabled_only?}, command_run {id, params?}, command_batch {steps, stop_on_error?}, doc_inspect {doc?}, and render_preview {doc?, page?, max_side?}; existing task tools (doc_open, doc_save, page_*, form_*, sign_*, ocr_*, …) remain available. pdfcraft://document and pdfcraft://commands resources mirror doc_inspect and command_list. Discover command ids and mapped tool schemas before running commands; unknown top-level MCP arguments are refused with JSON-RPC -32602, and a failed batch reports per-step errors. Inspect the returned doc id after file.open and pass it to dependent commands.
File access and output
Read and write roots are separate launch-time capabilities: every file a tool reads must lie inside the read root and every file it writes inside the write root; a root left out grants none of that access. Paths are resolved (links followed) and compared by whole folder names; a refused path is an ordinary, correctable tool error — do not bypass it with ambient filesystem commands. --root confines everything and can be combined with the automation roots (both must accept the path). Keychain or certificate-store signing identities and printer jobs are not file access and are not covered.
Exports complete synchronously; there is no background progress or cancellation. Save the requested deliverable (edited PDF in place or to a new path) and verify outputs exist and render: re-open or doc_inspect the saved file and render_preview the pages that matter, checking text, annotations, form fields, and page order. For batches, read the JSON report and surface skipped files and per-file errors.