Skip to content

MCP server

Doctopus ships a Model Context Protocol server, so agents can parse, chunk and compare documents in a folder you choose.

doctopus mcp --root ./docs

The server runs over stdio. Paths may be relative to --root or absolute within it; paths and symlinks that resolve outside the root are rejected.

Client configuration

{
  "mcpServers": {
    "doctopus": {
      "command": "doctopus",
      "args": ["mcp", "--root", "/path/to/docs"]
    }
  }
}

{
  "mcpServers": {
    "doctopus": {
      "command": "uvx",
      "args": ["--python", "3.12", "--from", "doctopus-ai", "doctopus", "mcp", "--root", "/path/to/docs"]
    }
  }
}

{
  "mcpServers": {
    "doctopus": {
      "command": "uv",
      "args": ["--directory", "/path/to/doctopus", "run", "doctopus", "mcp", "--root", "/path/to/docs"]
    }
  }
}

For Claude Code:

claude mcp add doctopus -- doctopus mcp --root /path/to/docs

Use --backend (or DOCTOPUS_BACKEND) to set the default engine for tool calls that do not name one.

Tools

Tool What it does
parse_document Returns Markdown, metadata, tables, and processing details.
chunk_document Returns paged chunks for retrieval or dataset building.
explain_routing Returns the routing decision for a file without extracting it.
list_processing_backends Lists engines with availability and capabilities.
compare_processors Runs every available engine with fallback disabled and reports per-engine results or failures.
  • parse_document caps Markdown at 60,000 characters by default (max_chars); truncated and total_chars describe the full length. Blocks are included only with include_blocks=True. It returns a subset of the canonical document, not its full JSON schema.
  • chunk_document pages its results with offset and limit (default 50), plus total. Each call parses the file again.
  • parse_document, chunk_document and explain_routing accept a backend argument.