Agentic Document Processing#

Docling exposes its document conversion capabilities to AI agents through three complementary integration paths: an MCP server (docling-mcp) for protocol-native agent integration, direct framework connectors for AG2/AutoGen, CrewAI, and Bee, and a proposed gRPC streaming layer for polyglot and low-latency use cases.

The shared output across all paths is the DoclingDocument JSON — a structured, richly-typed representation of converted documents that agents can reason over.

MCP Server (docling-mcp)#

The docling-mcp package implements a Model Context Protocol server that exposes Docling's conversion, generation, and RAG capabilities as MCP tools. It is a core runtime dependency of docling-serve 1.

Quick install for Claude Desktop or LM Studio — add the following to the client's config file 2:

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

Three installation profiles are available 1:

Transport protocols: stdio (Claude Desktop, LM Studio), sse (Llama Stack), and streamable-http (containerized/remote deployments) 1.

Remote API mode: To delegate conversion to a running docling-serve instance rather than running models locally 3:

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

Exposed tools: Conversion (PDF → DoclingDocument), generation (create DoclingDocument exportable to multiple formats), and RAG (Milvus upload/retrieval) 1. Examples using LlamaIndex, Llama Stack, Pydantic AI, and smolagents are in the docling-mcp examples directory 4.

AG2 / AutoGen Multi-Agent Integration#

Docling can be integrated directly into AG2 (AutoGen) multi-agent workflows by registering DocumentConverter operations as agent tools. The reference example is docs/examples/ag2_multiagent_document_analysis.ipynb.

Agent architecture in the example 5:

• document_processor — calls convert_document and extract_tables tools backed by Docling
• analyst — receives extracted content and produces a structured summary
• user_proxy — orchestrates the GroupChat and executes tool calls

Tool registration pattern — Docling's DocumentConverter is registered directly as an AG2 tool 6. The converter instance is stateless and thread-safe, so it can be shared across agents.

Install: pip install docling "ag2[openai]>=0.11.4,<1.0" pandas 7.

Other Framework Connectors#

• CrewAI — Docling is available as the CrewDoclingSource knowledge source. See docs/integrations/crewai.md.
• Bee Agent Framework — Docling is available as an extraction backend in Bee. See docs/integrations/bee.md.
• Agent skills file — docling/.agents/skills/docling/SKILL.md provides a decision guide for coding agents, covering CLI, Python SDK, DocumentExtractor, chunking/RAG, and remote service client use cases 8.

Proposed gRPC Streaming (Experimental, Not Merged)#

A community contributor has proposed a gRPC transport layer in two companion PRs 9:

• docling-serve PR #504 — gRPC server implementation and Python stubs
• docling-core PR #546 — protobuf definitions

Design principles 9:

• Pydantic-first: the Pydantic domain model stays the source of truth; protobuf is only the transport contract.
• Startup schema validation: the gRPC server cross-checks protobuf descriptors against the live Pydantic model at boot, failing fast on type drift.
• Semantic parity, not field-for-field REST mirroring: same document semantics and options, exposed through a gRPC-native API shape.

Planned server-side streaming would emit document parts incrementally (text blocks, tables, pictures) as parsing completes, rather than waiting for the full DoclingDocument. A draft streaming schema (StreamDocumentRequest / DocumentStreamEnvelope) using a oneof payload has been shared in the discussion thread 9.

Infrastructure caveat: Server-streaming gRPC requires a stable HTTP/2 connection. Environments that scale down replicas mid-request (e.g., Azure Container Apps GPU scaling) can sever long-lived streams — a known concern raised in the discussion 9.

Working client examples in Go, Java, Python, JavaScript, and Rust are available at ai-pipestream/docling-grpc-examples 9. The PR author has noted they will continue maintaining the gRPC fork even if the upstream PR is not merged 9.

Key References#