clichefactory-mcp

MCP (Model Context Protocol) server for ClicheFactory — structured data extraction from documents.

This server exposes ClicheFactory's extraction and document conversion capabilities as MCP tools, allowing AI assistants in Cursor, Claude Desktop, OpenClaw, and other MCP-compatible clients to extract structured data from PDFs, images, DOCX, XLSX, CSV, EML, and more.

Quick start (recommended — service mode)

Service mode uses the ClicheFactory cloud for the best extraction quality. You only need one API key.

1. Sign up at clichefactory.com — free pages included, no credit card required.

2. Create an API key in Settings → API Keys (format: cliche-...).

3. Install the MCP server:

   Terminal

   Copy

   pip install clichefactory-mcp
   

4. Configure — either paste the key into your MCP client (see below) or run once in a terminal:

   Terminal

   Copy

   pip install clichefactory   # if you don't have the CLI yet
   clichefactory configure
   

   The interactive wizard saves credentials to ~/.clichefactory/config.toml, which the MCP server reads automatically.

That's it — one env var (CLICHEFACTORY_API_KEY) or a config file, and you're on hosted extraction.

Tools

extract

The main tool. Pass a document file and a JSON schema — get structured data back.

Supports all extraction modes:

The schema can be provided as:

• File path: absolute path to a .json schema file
• Inline dict: the LLM constructs a JSON schema from the conversation (e.g., the user says "extract the invoice number and total" and the LLM builds {"type": "object", "properties": {"invoice_number": {"type": "string"}, "total": {"type": "number"}}})

to_markdown

Converts any supported document to markdown. Useful for inspecting document contents or feeding them to the LLM for analysis before deciding on an extraction schema.

doctor

Runs diagnostics on the ClicheFactory setup — config file, API keys, Python dependencies, system binaries. Call this when things aren't working.

Execution Modes

The server defaults to service mode (ClicheFactory cloud). Local mode is available for BYOK / air-gapped use.

• service (recommended) — Uses the ClicheFactory cloud service. Requires a ClicheFactory API key. Supports all extraction modes including trained pipelines and robust verification. Best extraction quality out of the box.

• local (advanced) — Runs extraction on your machine. You bring your own LLM key (BYOK). Requires pip install "clichefactory-mcp[local]" (~2 GB of parsing/OCR dependencies) plus system binaries (tesseract, LibreOffice). Quality depends on your local setup.

Installation

Prerequisites

• Python ≥ 3.12
• uv (recommended) or pip

From PyPI

pip install clichefactory-mcp

For local-mode extraction (BYOK, runs on your machine), install with the local extras:

pip install "clichefactory-mcp[local]"

Configuration

Environment Variables

Set these in your MCP client configuration (see below) or in ~/.clichefactory/config.toml via clichefactory configure.

Environment variables take precedence over the config file at ~/.clichefactory/config.toml.

Cursor

Add to .cursor/mcp.json in your project (or global Cursor settings):

{
  "mcpServers": {
    "clichefactory": {
      "command": "uvx",
      "args": ["clichefactory-mcp"],
      "env": {
        "CLICHEFACTORY_API_KEY": "cliche-your-key-here"
      }
    }
  }
}

For local development from a git checkout, replace uvx with:

"command": "uv",
"args": ["--directory", "/absolute/path/to/cliche-mcp", "run", "clichefactory-mcp"]

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "clichefactory": {
      "command": "uvx",
      "args": ["clichefactory-mcp"],
      "env": {
        "CLICHEFACTORY_API_KEY": "cliche-your-key-here"
      }
    }
  }
}

OpenClaw

Register the MCP server with your OpenClaw agent:

openclaw mcp set clichefactory '{"command":"uvx","args":["clichefactory-mcp"],"env":{"CLICHEFACTORY_API_KEY":"cliche-your-key-here"}}'

Verify with openclaw mcp list. The agent can now use extract, to_markdown, and doctor tools in any conversation.

An OpenClaw skill with agent instructions is also available in integrations/openclaw/. To install it into your workspace:

cp -r /path/to/cliche-mcp/integrations/openclaw ~/.openclaw/skills/clichefactory

Or, once published to ClawHub:

openclaw skills install clichefactory

Local mode (advanced)

If you prefer BYOK extraction on your machine, install the local extras and set LLM credentials:

{
  "mcpServers": {
    "clichefactory": {
      "command": "uvx",
      "args": ["clichefactory-mcp"],
      "env": {
        "LLM_MODEL_NAME": "gemini/gemini-3-flash-preview",
        "LLM_API_KEY": "your-gemini-api-key"
      }
    }
  }
}

Pass mode="local" explicitly in tool calls, or run clichefactory configure --local to set local as the default in ~/.clichefactory/config.toml.

Supported File Types

PDF, PNG, JPG, JPEG, WebP, GIF, BMP, DOCX, DOC, ODT, XLSX, CSV, EML, TXT, MD.

Differences from the CLI

This MCP server covers the core extraction and conversion workflows. The following CLI features are not included in v1:

Development

# Install in development mode
uv sync

# Run the server directly (stdio transport, for testing with MCP clients)
uv run clichefactory-mcp

# Inspect available tools (requires mcp CLI)
uv run mcp dev cliche_mcp/server.py

License

MIT — Copyright (c) 2026 Urban Susnik s.p.