file2markdown MCP server

Let Claude, Cursor, or any MCP client convert documents and web pages to Markdown mid-conversation — no tab switching. Paste one URL into your client and you're set.

Tools

• convert_url — fetch any public URL (web page, PDF, DOCX, and more) and get Markdown back. Web pages convert from served HTML; JavaScript-only content may be incomplete.
• convert_base64 — convert file contents directly, for programmatic clients.
• convert_batch — 2–10 documents in one call (Pro API key required). URLs and base64 files can be mixed; items convert in parallel and are isolated, so one bad URL returns an error in its slot without failing the rest. The output budget is shared across items; pass include_markdown: false to get only metadata and quality per item.
• list_supported_formats — formats, limits, and what needs Pro.
• usage_status — your tier and remaining conversions today.

All conversion tools return a quality field: an output-health check of the returned Markdown (status, coverage, and any flagged issues with line ranges — it does not compare against the source). Pass quality_check: "ai" to also send selected excerpts of the converted text — never your source file — to TypeSafe for narrow yes/no checks; it is off unless requested and works for PDFs only for now.

Why not just ask the model?

Because asking an LLM for “the markdown” of a document makes it retype the document — and models paraphrase, invent table cells, and silently truncate long files. This server runs real extraction engines instead:

• → Exact, not approximate. Engine-extracted text and tables, identical output every run — no hallucinated cells, no paraphrasing.
• → Files your model can't open. LLMs can fetch web pages, but not a DOCX, XLSX, PPTX, or EPUB behind a URL — and they can't OCR a scanned PDF at all.
• → Whole documents, cheap. A model transcribing 100 pages burns its context window and still won't finish. This returns clean Markdown up to 200k characters in one tool call.

For a short plain-text snippet, sure — just ask your model. The moment there's a table, a long document, a scan, or a file behind a URL, extraction beats regeneration.

Add to Claude

claude.ai: Settings → Connectors → Add custom connector → paste the endpoint URL. When asked about authentication, choose None — this server uses API keys, not OAuth. To use a Pro key, add a request header in the same dialog: name Authorization, value Bearer f2m_…. Without a key you get the free tier (5 conversions/day per network).

Claude Code:

claude mcp add --transport http file2markdown https://mcp.file2markdown.ai/mcp

Cursor / generic clients (JSON config):

{
  "mcpServers": {
    "file2markdown": {
      "url": "https://mcp.file2markdown.ai/mcp"
    }
  }
}

Tiers

The free tier is limited per network IP, so shared networks (and hosted agents) may hit it sooner. A Pro key removes the limit: create one on your account page and send it as Authorization: Bearer f2m_…. Plans: pricing.

In-browser tools (WebMCP)

file2markdown's pages also register WebMCP tools for browser-integrated agents (Chrome 149+ origin trial, Edge 150, ChatGPT Desktop). On the convert page an agent can run conversions on files you added — the one thing a remote server can't reach is your local disk — plus convert_website and format_for_linkedin on the free tools. Your signed-in session carries your plan: no API key needed in the browser.

Prefer a browser?

The same converter runs at /tools/website-to-markdown for URLs, and on the main converter for file uploads.