markitdown
/markitdown
Zamienia pliki PDF, Office, HTML, obrazy, audio i adresy URL na Markdown, aby wyciągnąć tekst lub tabele.
- Autor
- AlbertZJ ↗
- Repozytorium
- dsh-markitdown ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 26.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 26.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Zamienia pliki PDF, Office, HTML, obrazy, audio i adresy URL na Markdown, aby wyciągnąć tekst lub tabele.
Oryginalny opis (EN)
Convert PDF, Office, HTML, image, audio and other files, URLs, or stdin streams to Markdown with the dsh-markitdown plugin (Microsoft MarkItDown). Use when the user asks to extract text or tables from a document, convert a file or YouTube link to Markdown, or make an attachment readable.
Jak działa
- Workflow
- Hints and flags
- URL sources and egress
- DOCX images: export by default, descriptions optional
- Supported local inputs
- Rules
- Known limitations (set expectations)
Podgląd SKILL.md
---
name: markitdown
description: Convert PDF, Office, HTML, image, audio and other files, URLs, or stdin streams to Markdown with the dsh-markitdown plugin (Microsoft MarkItDown). Use when the user asks to extract text or tables from a document, convert a file or YouTube link to Markdown, or make an attachment readable.
---
# MarkItDown document conversion
Convert sources to Markdown through the plugin's `markitdown_check` and
`markitdown_convert` tools. MarkItDown runs as a Python package
(`python -m markitdown`) in the plugin's configured interpreter; the tools
own the process, timeout and error handling.
## Workflow
1. **Verify the environment once per session** with `markitdown_check`.
- `ready: true` → continue.
- `ready: false` → the detail field contains the exact setup command
(`<command> -m pip install 'markitdown[all]'`). Show it to the user and
stop; do not retry conversions until they confirm setup.
2. **Convert** with `markitdown_convert`. Pick the mode:
- `path`: one absolute local file path or an http(s) URL (e.g. a YouTube
link, when the `[all]` extra is installed).
- `paths`: a batch (up to 16) of paths/URLs; per-source failures are
captured in `results` and do not abort the batch.
- `stdinBase64`: canonical base64 bytes; requires `extensionHint` or
`mimeHint` so MarkItDown can pick a converter.
3. **Where the Markdown goes**:
- Omit `output`/`outputDir` to get Markdown inline — right for summaries
and sources under ~100 KiB.
- `output` (single source) or `outputDir` (writes `<name>.md` per source)
— right for large documents, batch jobs and when the user asks for
converted files on disk. DOCX conversions also write a sibling
`<name>.media/` directory with the embedded images.
4. **Report** what was converted, where the Markdown went, and any caveats
below that apply to the source format.
## Hints and flags
`extensionHint` / `mimeHint` / `charsetHint` map to the MarkItDown CLI's
`-x` / `-m` / `-c` flags. `extensionHint` also admits local files whose
extension is not in the built-in supported list, and is required for stdin;
hinted extensions are validated against the supported list plus a small set
of plain-text formats (txt, log, yaml, yml, toml, ini, cfg, conf, rss, atom,
py, js, ts, css, sh). `keepDataUris` keeps base64 image payloads (by default
they are truncated). Third-party MarkItDown converter plugins (`-p`) run
only when the plugin admin enables `allowPlugins` in the plugin config.
## URL sources and egress
`path`/`paths` accept http(s) URLs (e.g. YouTube links). Every URL is
downloaded with connection-pinned egress validation before conversion: the
plugin resolves the host itself, connects only to a globally routable
address (the checked address is the connected address, so DNS rebinding
cannot swap targets mid-handshake), follows redirects manually with the
same check per hop, and hands the downloaded bytes to MarkItDown via a
temporary file — MarkItDown never re-resolves the host. Internal addresses
(loopback, private, link-local) are refused unless the admin sets
`allowPrivateUrls`. YouTube transcript URLs are the documented exception:
the URL itself is passed to MarkItDown (its converter requires it), still
after the same pinned connection check.
## DOCX images: export by default, descriptions optional
Markitdown 0.1.x never routes Word images through its LLM client, so a DOCX
would otherwise carry `` blobs. The bundled
helper instead exports every `word/media` image in document order next to the
Markdown (`<name>.media/`) and replaces each placeholder with a file link —
no model, no API key. Inline conversions use a temp directory (absolute
links); pass `output`/`outputDir` when the images should persist.
When the user wants the text inside the screenshots without opening each
image, the model can describe them: enable **Describe images with the active
model** on Settings → MarkItDown (`describeImages`; off by default). Each link
then carries a `> **图 N**(word/media/...)` description block. This costs one
model call per distinct image and needs:
- the `openai` package in the same Python environment
(`<command> -m pip install openai`),
- an active provider with its API key stored in the dsh credential store —
Settings → MarkItDown (endpoint, model, key) — or exported in the
`llmApiKeyEnv` environment variable as a fallback.
Descriptions run at temperature 0 and must transcribe text verbatim, mark
unreadable text and undeterminable senders instead of guessing, and name the
source part so the original can be inspected. Large images (> 8 MiB) are not
exported or described (reported inline), and a failed description is reported
inline instead of aborting the document. Descriptions are still one-shot:
when exact wording matters, open the exported image (the session's own image
reading) instead of trusting the text.
`markitdown_check` reports whether descriptions are on and whether the helper
can reach a key. Without them, conversion stays fully offline and images are
only exported.
## Supported local inputs
PDF, DOCX/DOC, PPTX/PPT, XLSX/XLS/CSV (tables become Markdown tables),
JSON/XML, HTML/HTM, EPUB, MD, Outlook MSG, ZIP (each contained file
converted in place), JPG/JPEG/PNG/GIF/TIFF/BMP (EXIF metadata, OCR text
when available), WAV/MP3 (speech transcription, needs the `[all]` extra).
Anything else: refuse conversion and say the format is not supported,
unless the user provides an `extensionHint` that MarkItDown understands.
## Rules
- Always run `markitdown_check` before the first conversion in a session;
a green check earlier in the conversation counts.
- Local paths must be absolute and the file must exist; URLs must start
with http(s)://. Errors say which check failed.
- Files above the configured size limit are rejected — suggest splitting
or shrinking the file.
- Batch jobs run sequentially inside the tool; never fire several
`markitdown_convert` calls in parallel for the same batch.
- Do not install Python packages yourself unless the user explicitly asks;
surface the setup command from `markitdown_check` and let them decide.
## Known limitations (set expectations)
- Scanned PDFs without a text layer convert to little or no text; say so
instead of presenting an empty result as success.
- Complex layouts (multi-column, heavy styling) flatten into linear Markdown.
- YouTube transcription depends on the `[all]` extra and current page
structure; failures surface the MarkItDown error — report it, do not retry
in a loop.
- DOCX embedded images are exported as files beside the Markdown. Inline
conversions use a temp directory that may be cleaned up; prefer
`output`/`outputDir` when the images matter. With descriptions enabled,
expect one model call per distinct image and scale `timeoutMs` for
image-heavy documents.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · AlbertZJ/dsh-markitdown | 90 | MIT |
| TerminalSkills/skills | 100 | Apache-2.0 |
| pleaseai/claude-code-plugins | 100 | MIT |
| matematicsolutions/awesome-matematic-skills-pl | 100 | MIT |
| rashadansari/myagents | 100 | MIT |
| julianobarbosa/claude-code-skills | 100 | MIT |
| jacardl/viceroy-skills | 98 | MIT |
| 0xDarkMatter/claude-mods | 95 | MIT |
| jackspace/ClaudeSkillz | 95 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki d59bc9707046 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100markitdown-skill
Konwertuje dokumenty i adresy URL do Markdown poleceniem markitdown: PDF, Word, PowerPoint, Excel, obrazy z OCR, audio z transkrypcją, HTML i YouTube, także wsadowo. - 100convert-to-markdown
Konwertuje lokalne pliki PDF, Office, arkusze, prezentacje, HTML i CSV na czysty Markdown przed czytaniem, używając MarkItDown, co ogranicza zużycie tokenów. - 100alterlab-markitdown
Konwertuje pliki i dokumenty Office na czysty Markdown przyjazny dla LLM za pomocą Microsoft MarkItDown: PDF, DOCX, PPTX, XLSX, obrazy, audio, HTML, EPUB i inne, z OCR. - 100markitdown-convert
Konwertuje pliki do Markdown przy użyciu Microsoft MarkItDown, m.in. PDF, Word, Excel, PowerPoint, EPUB, obrazy, audio, HTML, archiwa ZIP, notebooki Jupyter i pliki .msg z Outlooka. - 100markdown-converter
Konwertuje PDF, Word, PowerPoint, Excel, HTML, CSV, JSON, obrazy z OCR, audio z transkrypcją, archiwa ZIP, adresy YouTube i EPUB do Markdown za pomocą markitdown. - 100to-markdown
Zamienia pliki i adresy URL, np. PDF, DOCX, XLSX, HTML, obrazy i nagrania, na Markdown dla potoków LLM.