localparse-mcp

An MCP server that gives your coding agent (Cursor, Claude Desktop, etc.) accurate document parsing and schema-driven JSON extraction via LocalParse — table-detection recovery + arithmetic identity checks on the parse side, and a server-validated JSON Schema contract on the extract side — instead of letting it write throwaway OCR scripts.

Tools

Parse

result_type ∈ markdown (default), text, json, structured.

Extract (Beta)

The server validates the model's output against your data_schema before returning, so the JSON you get back is guaranteed to match the shape you asked for. Example natural-language prompts an agent can use:

"Use extract_document on ./invoices/q3.pdf with this schema: {type:'object', properties:{vendor:{type:'string'}, total:{type:'number'}}, required:['vendor','total']}"

Configure your client

You need a LocalParse API key (LOCALPARSE_API_KEY).

Cursor — ~/.cursor/mcp.json (or project .cursor/mcp.json)

{
  "mcpServers": {
    "localparse": {
      "command": "uvx",
      "args": ["localparse-mcp"],
      "env": { "LOCALPARSE_API_KEY": "lp-your-key" }
    }
  }
}

Claude Desktop — claude_desktop_config.json

{
  "mcpServers": {
    "localparse": {
      "command": "uvx",
      "args": ["localparse-mcp"],
      "env": { "LOCALPARSE_API_KEY": "lp-your-key" }
    }
  }
}

Then ask the agent things like "parse the PDFs in ./data-room into case acme and tell me which tables don't reconcile."

Local development (before it's on PyPI)

Run the server straight from the repo and point the client at it:

{
  "mcpServers": {
    "localparse": {
      "command": "uv",
      "args": ["run", "--directory", "/abs/path/to/Agent-ingestor", "localparse-mcp"],
      "env": { "LOCALPARSE_API_KEY": "lp-your-key" }
    }
  }
}

Point at a self-hosted parser by also setting LOCALPARSE_BASE_URL.