MarkItDown Claude Skill: How It Works, and When to Use an MCP Server Instead
MarkItDown Claude Skill: How It Works, and When to Use an MCP Server Instead
Searching for a "MarkItDown Claude Skill" turns up a handful of community-built SKILL.md files on skill marketplaces, but no standalone MarkItDown Skill from Microsoft or Anthropic. Claude's pre-built Skills don't include a generic "convert any file to Markdown" capability, so builders have filled the gap themselves. Before installing one, it's worth understanding what a Skill actually is, how it differs from an MCP server, and what Anthropic's own docs say about running third-party ones. If you just want files converted to Markdown inside Claude right now, file2markdown runs a hosted MCP server that does that with nothing to install.
What a Claude Agent Skill Actually Is
Agent Skills are a specific mechanism Anthropic ships across claude.ai, the Claude API, and Claude Code: a folder containing a SKILL.md file with YAML frontmatter (name and description) plus, optionally, extra reference files and scripts. Claude loads the name/description pair into its system prompt at startup (cheap: about 100 tokens per Skill, per Anthropic's docs) and only reads the full SKILL.md body (and any bundled files) when a request actually matches that description. Anthropic calls this progressive disclosure: a Skill can bundle a large reference doc or a utility script without any context cost until Claude actually needs it.
That's a different mechanism from an MCP server. An MCP server exposes callable tools over a protocol: Claude sends a request, the server runs code and returns a result. A Skill is closer to a reusable onboarding doc: it's instructions (and optionally scripts) that live on a filesystem Claude can read and execute with bash. Anthropic's Agent Skills docs are explicit that Skills require a code-execution environment: on the Claude API that means the code execution tool; in Claude Code it's the local filesystem.
Anthropic's own pre-built Skills are PowerPoint, Excel, Word, and PDF, and they're for creating, editing, and analyzing those document types, not for converting arbitrary files into Markdown. MarkItDown does show up inside Anthropic's work, though: in its open-source skills repository, the pptx and xlsx Skills call the markitdown command to read decks and spreadsheets. What doesn't exist is an official standalone "markitdown" or "any-to-markdown" Skill, and Microsoft's markitdown repository ships a Python library, a CLI, and an MCP server rather than a Skill. The community has filled that gap with Skills, listed on marketplaces like Smithery, LobeHub, and Skills Directory, that wrap the markitdown library in a SKILL.md.
The Catch: Network Access Depends on Where You Run It
Before installing a community MarkItDown Skill, it matters which Claude surface you're using, because the runtime constraints are genuinely different:
- Claude API: Skills run in a sandboxed container with no network access and no runtime package installation: only pre-installed packages are available, and markitdown isn't on the code execution tool's documented list of pre-installed libraries. A Skill that assumes it can
pip install markitdownon demand, or fetch a file from a URL, simply can't do either there. - Claude Code: Skills have the same network access as any other program on your machine, and live as plain directories: personal Skills in
~/.claude/skills/, project Skills in.claude/skills/. A markitdown-wrapping Skill here can install dependencies and read local files with your own user privileges. - claude.ai: Custom Skills are uploaded as ZIP files under Customize → Skills, and network access varies by user/admin settings. An uploaded Skill stays private to your account until you share it; on Team and Enterprise plans you can share it with colleagues, and organization owners can provision Skills for everyone.
That point about Claude Code (a Skill running with your file privileges) is the same tradeoff we cover for MarkItDown's MCP server: a local tool that reads files on your behalf is only as safe as its code and its scope. Anthropic's security guidance for Skills is blunt about this: use Skills only from sources you trust (yourself, or Anthropic), and if you use a third-party one, audit SKILL.md and any bundled scripts first, because a Skill can direct Claude to run tools or code that doesn't match its stated purpose. The standalone markitdown Skills you'll find today are community-maintained, not from Microsoft or Anthropic, so that guidance applies directly before you add one to a project with real file access.
A No-Install Alternative: file2markdown's MCP Server
If what you actually want is "convert this PDF/DOCX/PPTX to Markdown without leaving my Claude conversation," an MCP server gets there without a SKILL.md to vet. file2markdown runs a hosted one at mcp.file2markdown.ai/mcp:
claude.ai: Customize → Connectors → + Add → Add custom connector, name it, paste the endpoint URL, and choose "No sign in" under Authentication (it uses API keys, not OAuth). To use a Pro key, add a request header in the same dialog: Authorization: Bearer f2m_….
Claude Code:
claude mcp add --transport http file2markdown https://mcp.file2markdown.ai/mcp
Cursor or any JSON-config client:
{
"mcpServers": {
"file2markdown": {
"url": "https://mcp.file2markdown.ai/mcp"
}
}
}
It exposes convert_url and convert_base64 tools that hand back clean Markdown for PDF, DOCX, PPTX, XLSX, HTML, and more, with headings, lists, and tables preserved instead of an LLM retyping (and subtly paraphrasing) the document from a pasted copy. The free tier needs no signup: 5 conversions a day per network, files up to 10 MB sent as base64 or 25 MB fetched from a URL, and no OCR. A Pro key removes the daily cap, raises the limit to 100 MB, and adds OCR for scanned PDFs and images. The server never touches your local filesystem: it only sees the file contents or URL your client sends it, converts them on file2markdown's servers, and deletes the file right after (image and scanned-PDF OCR passes the file to Anthropic's Claude). There's no SKILL.md, no bundled script, and no local package install to audit.
Skill vs. MCP Server: The Practical Difference
| Claude Skill | MCP Server | |
|---|---|---|
| What it is | A folder with SKILL.md + optional scripts, read via bash | A networked tool Claude calls over a protocol |
| Where it runs | Claude's code-execution environment (varies by surface) | Wherever the server is hosted |
| Network access | None on the Claude API; full in Claude Code; varies on claude.ai | Defined by the server itself |
| Trust model | You're trusting the Skill's instructions and any bundled code | You're trusting the server's responses to defined tool calls |
| Setup for file2markdown | Not applicable: there is no file2markdown Skill | Add one URL, works today |
If you're building a custom Skill for your own workflow (summarizing a specific internal doc format, say), Skills are a good fit; that's what they're designed for. For "turn documents into Markdown," an MCP server already does the job without asking you to trust and maintain a third-party SKILL.md.
Frequently Asked Questions
Is there an official MarkItDown Skill from Microsoft or Anthropic?
Not a standalone one. Anthropic's pre-built Agent Skills are PowerPoint, Excel, Word, and PDF. Its open-source skills repository publishes those document Skills plus others, such as the Claude API skill and skill-creator, and its pptx and xlsx Skills use the markitdown command to read files, but none of them is a general file-to-Markdown converter. Microsoft's markitdown repository has no Skill either. The MarkItDown Skills you'll find on marketplaces like Smithery and LobeHub are community-built, not official.
What's the real difference between a Claude Skill and an MCP server?
A Skill is instructions (and optional scripts) that live on a filesystem Claude reads and executes with bash; it's closer to a reusable playbook. An MCP server is a separate process exposing callable tools over the Model Context Protocol. Skills need a code-execution environment to run; MCP servers just need a client that can reach their endpoint.
Can I use a document-conversion Skill on the Claude API?
Only if it doesn't need network access or packages beyond what's pre-installed: the Claude API's code-execution sandbox has no internet access and no runtime pip install, and markitdown isn't on its documented list of pre-installed libraries. A Skill that assumes it can fetch a URL or install markitdown on the fly will fail there, even though the same Skill works fine in Claude Code, which has full network access.
Is it safe to install a community MarkItDown Skill?
That depends entirely on the specific SKILL.md and any scripts it bundles. Anthropic's own guidance is to audit third-party Skills before use, since they can direct Claude to run code with whatever access it has. If you'd rather not review someone else's SKILL.md, file2markdown's MCP server converts PDF, Word, PowerPoint, Excel, and the other formats it accepts without installing anything locally. It accepts fewer inputs than MarkItDown, which also reads things like YouTube URLs.
Want Markdown out of Claude right now instead of auditing a Skill first? Add file2markdown's MCP server or convert a file directly, with no install either way.
The Markdown Memo
A fortnightly note for lawyers, researchers, accountants, and anyone else drowning in PDFs, scans, and decks. No spam.