Md Anything

md-anything

Convert files, URLs, and media into Markdown for agent workflows and local automation.

Install: bun install -g md-anything or npm install -g md-anything

Quick commands

convert

Convert a single file or URL to Markdown.

mda <input>
mda convert <input>
mda convert <input> -o output.md
mda convert "https://example.com/article"
mda convert report.pdf
mda convert image.png
mda convert video.mp4
mda convert "https://www.youtube.com/watch?v=..."

ingest

Batch-convert all supported files in a folder.

mda ingest ./notes
mda ingest ./notes -o ./output
mda ingest ./vault -r -o ./output

doctor / examples

Report capabilities or see copy-paste examples.

mda doctor
mda examples
mda demo

[!NOTE] mda mcp install assumes md-anything-mcp is in your PATH. For binary-only installs, use the manual config shown below with bunx.

Options

Supported Input Types

Capability Layers

md-anything is designed to stay lightweight by default:

• Core mode: useful with zero extra native tools
• Local upgrade mode: add tesseract, pdftotext, whisper-cpp, ffmpeg, or ebook-convert as needed
• Remote enhancement mode: set OPENROUTER_API_KEY only if you explicitly want remote image/audio/video fallbacks

No local models are bundled in the package. Heavy media capabilities remain optional layers, not requirements.

JSON Output

Use --json to get structured output consumable in agent workflows:

mda convert report.pdf --json

{
  "input": "report.pdf",
  "markdown": "# Report Title\n...",
  "kind": "pdf",
  "supportLevel": "strong",
  "chunks": [],
  "metadata": {
    "title": "Report Title",
    "source": "report.pdf",
    "extraction": "unpdf",
    "extraction_status": "ok",
    "support_level": "strong",
    "usefulness_score": 0.85
  },
  "provenance": { "documentId": "..." },
  "warnings": []
}

For ingest:

mda ingest ./notes --json

{
  "converted": 12,
  "skipped": 2,
  "failed": 0,
  "docs": [{ "fileName": "note.md", "title": "My Note", "sourceType": "pdf", "source": "report.pdf", "metadata": { "extraction_status": "ok" } }]
}

Weak or uncertain results still return JSON, with warnings instead of silent failure:

{
  "input": "unknown-input.foo",
  "markdown": "# ...",
  "kind": "unknown",
  "supportLevel": "optional",
  "metadata": {
    "extraction_status": "weak"
  },
  "warnings": [
    "Could not detect a supported input type for \"unknown-input.foo\".",
    "Try `mda --help` or `mda examples` for supported usage."
  ]
}

Argument errors are also machine-readable:

{
  "error": "Missing input for convert command.",
  "code": "missing_input",
  "examples": [
    "mda convert tests/fixtures/sample.txt",
    "mda convert \"https://example.com/article\""
  ]
}

MCP Server

Add to .mcp.json for use inside Claude, Cursor, or other MCP hosts:

{
  "mcpServers": {
    "md-anything": {
      "command": "bunx",
      "args": ["md-anything-mcp"]
    }
  }
}

The MCP server exposes:

• tools: convert, ingest, doctor
• resources: md-anything://doctor, md-anything://workspace-policy, md-anything://workspace/{path}
• prompts: analyze_document, summarize_document_chunks

Safety defaults for MCP:

• local paths must stay inside the current workspace
• only http:// and https:// URLs are allowed
• private and localhost URLs are blocked unless MDA_MCP_ALLOW_PRIVATE_URLS=1

Ecosystem Compatibility

md-anything is intentionally designed to work well in agent ecosystems that need:

• a stable CLI entrypoint
• machine-readable JSON output
• package-shipped skill metadata
• local execution with optional capability upgrades

That makes it a good fit for package-scan discovery systems, Claude Code plugin flows, and ClawHub/OpenClaw-style skill ecosystems.

Source: ojspace/md-anything — distributed by TomeVault.