any-extractor

One extract() call. Any document. Agent-ready markdown, typed blocks, and metadata.

Point it at a file, URL, or Buffer. Get back:

• markdown — a single GFM string, ready for an LLM.
• text — plain reading-order text, ready for embeddings or search.
• sections — ordered pages / slides / sheets with typed blocks.
• metadata — MIME, title, author, page count, sheet names.

import { extract } from 'any-extractor';

const result = await extract('./quarterly-report.pdf');

result.markdown; // GFM string
result.text; // plain text
result.sections; // typed blocks
result.metadata; // { mime, title, pageCount, ... }

result.markdown

# Q3 Results

Revenue grew **18%** year-over-year, driven by APAC.

| Region | Revenue |
| ------ | ------- |
| APAC   | $4.2M   |
| EMEA   | $3.1M   |

result.sections

[
  {
    kind: 'page',
    label: 'Page 1',
    index: 1,
    blocks: [
      { id: 'a1b2…', type: 'heading', level: 1, text: 'Q3 Results' },
      {
        id: 'c3d4…',
        type: 'paragraph',
        text: 'Revenue grew **18%** year-over-year, driven by APAC.',
      },
      {
        id: 'e5f6…',
        type: 'table',
        headers: ['Region', 'Revenue'],
        rows: [
          ['APAC', '$4.2M'],
          ['EMEA', '$3.1M'],
        ],
      },
    ],
  },
];

result.metadata

{
  mime: 'application/pdf',
  source: './quarterly-report.pdf',
  title: 'Q3 Results',
  author: 'Finance Team',
  pageCount: 42,
}

Install

npm install any-extractor

Node.js ≥ 18.

MCP Server

any-extractor doubles as a Model Context Protocol server. Drop it into Claude Desktop, Cursor, VS Code, Continue, or any MCP-capable agent.

{
  "mcpServers": {
    "any-extractor": {
      "command": "npx",
      "args": ["-y", "any-extractor-mcp"],
    },
  },
}

| Tool | Use it for | | ----------------------------- | ------------------------------------------------------------- | | extract_document | Default. Markdown + metadata + section index. | | extract_document_structured | Full typed section/block tree — for agents that walk content. | | extract_section | One section by index. Cheap paging for large PDFs. |

Supported formats

| Format | Sections emitted | | ------------ | -------------------------- | | PDF | one page per page | | Word | single body | | Excel | one sheet per worksheet | | PowerPoint | one slide per slide | | OpenDocument | body / sheet / slide | | HTML | single body | | Markdown | single body | | Plain text | single body | | CSV | single body | | JSON | single body |

CLI

# Markdown to stdout (default)
npx any-extractor report.pdf

# Everything else — flags, formats, URLs, stdin, timeouts
npx any-extractor --help

Programmatic API

Cancellation & timeouts

// User-driven cancel
const ac = new AbortController();
await extract('./big.pdf', { signal: ac.signal });

// Hard deadline
await extract(url, { timeoutMs: 10_000 });

// Both — whichever fires first wins
await extract(url, { signal: ac.signal, timeoutMs: 30_000 });

Custom parsers

Register your own MIME handler — e.g. route images through a vision LLM. User parsers override built-ins, and embedded images inside Word / PowerPoint / OpenDocument get enriched automatically.

import { AnyExtractor } from 'any-extractor';

const extractor = new AnyExtractor();

extractor.addParser({
  mimes: ['image/png', 'image/jpeg'],
  concurrency: 2, // rate-limit in-flight calls
  async parse(buffer, ctx) {
    const caption = await myVisionModel(buffer);
    return {
      sections: [{ kind: 'body', blocks: [ctx.block.paragraph(caption)] }],
    };
  },
});

await extractor.extract('./slides.pptx');

Enriched images render with a blockquote caption in the output markdown:

![Sales chart](media/image1.png)

> Bar chart showing Q3 revenue up 18% vs. Q2, driven by APAC.

Support

If any-extractor saved you an afternoon, you can — it keeps the parsers fed.

License

MIT