MCP Server Integration

MCP Server Integration#

Model Context Protocol (MCP) is the standard Docling uses to expose document processing capabilities to AI agents and LLM clients. The implementation lives in the dedicated docling-mcp package (currently v2.1.0), which is a core runtime dependency of docling-serve .

Architecture and Installation Modes#

Docling MCP v2.0 introduced a hybrid architecture with three installation profiles :

Mode	Install	When to use
Remote (default)	`pip install docling-mcp`	Access to a Docling Serve API; ~50 MB, no model downloads
Local	`pip install docling-mcp[local]`	Offline or no Docling Serve available; full model stack
Hybrid	`pip install docling-mcp[local]` + `DOCLING_FALLBACK_TO_LOCAL=true`	Remote preferred, automatic local fallback

For remote/hybrid modes, set these environment variables :

export DOCLING_SERVICE_URL=https://your-docling-service.example.com
export DOCLING_SERVICE_API_KEY=your-api-key # if required
export DOCLING_CONVERSION_MODE=remote

Transport Protocols#

The docling-mcp-server CLI supports three transports :

stdio — for Claude Desktop and LM Studio (default for local client installs)
sse — for Llama Stack integrations
streamable-http — for containerized/remote deployments

MCP Client Configuration#

The fastest way to wire up any MCP-compatible client is via uvx :

{
  "mcpServers": {
    "docling": {
      "command": "uvx",
      "args": ["--from=docling-mcp", "docling-mcp-server"]
    }
  }
}

For Claude Desktop, edit claude_desktop_config.json with the snippet above. For LM Studio, edit mcp.json. Example configs are in docs/integrations/.

Container Deployment via `docling-serve`#

The docling-serve container image (quay.io/docling-project/docling-serve) bundles the docling-mcp-server executable starting from v1.1.0 . By default, the container runs docling-serve run on port 5001. To launch the MCP server instead, override the entrypoint :

podman run -p 8000:8000 quay.io/docling-project/docling-serve \
  -- docling-mcp-server --transport streamable-http --port 8000 --host 0.0.0.0

The MCP endpoint is then available at http://localhost:8000/mcp. Clients connect to it via :

{
  "mcpServers": {
    "docling": { "url": "http://localhost:8000/mcp" }
  }
}

Exposed Tools#

The MCP server exposes tools for :

Conversion — PDF → DoclingDocument (structured JSON)
Generation — create DoclingDocument objects exportable to multiple formats
RAG — Milvus upload and retrieval integration

Framework-specific examples using LlamaIndex, Llama Stack, Pydantic AI, and smolagents are in the examples/ directory .

Key References#

docling-mcp repository — source, examples, migration guide
docling MCP usage docs — client setup, remote API mode, env vars
docling-serve MCP docs — container entrypoint override, port mapping
docling-serve pyproject.toml — docling-mcp>=1.0.0 dependency pin