Multi-Language API Integration#

Docling's core processing engine is Python-only, but it exposes document conversion capabilities to every other language ecosystem through Docling Serve — a FastAPI service that listens on port 5001 1. There are four integration paths:

The shared wire format across all paths is the DoclingDocument JSON schema 2.

HTTP REST API#

The REST API is the primary integration point for any language. Key endpoints 3:

A minimal call from any language reduces to a POST with a JSON body 4:

curl -X POST http://localhost:5001/v1/convert/source \
  -H 'Content-Type: application/json' \
  -d '{"sources": [{"kind": "http", "url": "https://arxiv.org/pdf/2501.17887"}]}'

The response contains md_content, json_content, html_content, etc., controlled by to_formats in the request options 5. Interactive API docs are at http://localhost:5001/docs 6.

When API key authentication is enabled, all requests must include the X-Api-Key header 7.

Java Integration (docling-java)#

docling-java provides two layers of Java API 8:

• docling-core (ai.docling:docling-core) — Jackson-annotated Java model mirroring the DoclingDocument JSON schema
• docling-serve-api (ai.docling:docling-serve-api) — framework-agnostic interface (DoclingServeApi) and request/response models
• docling-serve-client (ai.docling:docling-serve-client) — reference implementation using Java's built-in HttpClient and Jackson; resolves between Jackson 2/3 at runtime via SPI 9
• docling-testcontainers (ai.docling:docling-testcontainers) — official Testcontainers module for spinning up Docling Serve in JUnit tests

Quick start — see README for a full example using DoclingServeApi.builder(), ConvertDocumentRequest, and InBodyConvertDocumentResponse.

All artifacts are published to Maven Central and tested weekly against all published Docling Serve versions 9.

Framework integrations — the standalone client targets plain Java. For framework-managed contexts, use dedicated extensions 9:

Full Java client documentation: docling-project.github.io/docling-java.

Node.js / JavaScript: Bee Agent Framework#

Docling is available as a document extraction backend in the Bee Agent Framework — an open-source AI agent framework from IBM 10.

• NPM package: bee-agent-framework
• Docs: i-am-bee.github.io/bee-agent-framework

This integration allows Node.js/TypeScript agents to call Docling's document extraction capabilities without writing a REST client from scratch. The Bee framework handles the HTTP communication with a running Docling Serve instance.

Experimental gRPC Bindings#

Status: Not merged. PR #504 ("feat: Grpc native converter") has been open since February 2026, with active discussion as of July 2026. Do not use in production.

A community contributor has proposed a gRPC transport layer for Docling Serve in two companion PRs 11:

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

Design#

The approach is semantic parity, not a strict REST-to-gRPC endpoint mirror 11. Key properties:

• Pydantic-first: the Pydantic domain model remains the source of truth; protobuf is the transport contract
• Startup schema validation: the gRPC server cross-checks protobuf descriptors against the live Pydantic model at startup, failing fast on type drift
• Three proto namespaces under ai/docling/serve/v1/: docling_serve.proto, docling_serve_stream.proto, docling_serve_types.proto

Polyglot client examples#

The author provides working client examples in Go, Java, Python, JavaScript, and Rust at ai-pipestream/docling-grpc-examples 11.

Planned: server-side streaming#

The proposed streaming extension would emit document parts incrementally (text blocks, tables, pictures) as they complete parsing, rather than returning the full DoclingDocument at the end. A draft streaming protobuf schema (StreamDocumentRequest / DocumentStreamEnvelope) has been shared in the discussion thread 11.

Infrastructure note: pure server-streaming gRPC requires stable HTTP/2 connections. In environments that scale down replicas mid-request (e.g., Azure Container Apps GPU scaling), a long-lived gRPC stream can be severed — the same constraint that affects WebSocket-based async polling 11.