Use it from Claude, Cursor and other assistants (MCP)¶
readanydoc mcp runs ReadAnyDoc as an MCP server, so an AI assistant can read
your documents directly in a chat, without code.
It adds two tools that work on files on your computer:
| Tool | What it returns |
|---|---|
parse_document(path, mode?, pages?) |
the document as Markdown (tables as HTML), page count, warnings, cost |
extract_fields(path, schema, instructions?, array_extract?, mode?) |
values for a JSON Schema, each with citations (text, page box, confidence), warnings, cost |
Install¶
The server uses your own ANTHROPIC_API_KEY (or OPENAI_API_KEY for OpenAI models), passed in its environment.
Claude Code¶
The plugin adds the server and a skill that tells Claude when to use it. It needs uv and
ANTHROPIC_API_KEY in the environment you start Claude Code from:
Or add only the server:
Claude Desktop, Cursor, Windsurf and other clients¶
Add this to the client's MCP configuration (claude_desktop_config.json, .cursor/mcp.json, ...). With uv installed nothing else is needed; or pip install "readanydoc[engine,mcp]" and use
"command": "readanydoc", "args": ["mcp"] (the full path from which readanydoc if the client does not see your
virtual environment).
{
"mcpServers": {
"readanydoc": {
"command": "uvx",
"args": ["--from", "readanydoc[engine,mcp]", "readanydoc", "mcp"],
"env": {"ANTHROPIC_API_KEY": "your-key"}
}
}
}
Try it¶
Ask the assistant, for example: "Extract the invoice number, date and total from ~/Downloads/invoice.pdf with citations" or "Parse ~/Documents/report.pdf and summarise the tables on page 3".
Also listed in¶
The official MCP Registry as io.github.arpitpatel25/readanydoc.
Notes¶
- Files are read from the local disk only; the server does not fetch URLs.
- A parse can take from a few seconds to minutes on long documents.
mode="model-only"is faster. - Check
warningsin every result. Confidence ishighorlow, not a probability.
If ReadAnyDoc saved you time, a star on GitHub helps other developers find it.