Analytics cookies only (Google Analytics, PostHog). No ad networks, no cross-site tracking. See our privacy policy.
DS
Self-host Docling Serve
updated Nov 2026deployed by usprices checked · Jul 2026Install verified on Ubuntu 26.04 · Nov 2026 · how we test
We earn commissions when you shop through the links below. Full disclosure →
An HTTP API and small web playground for Docling, a document-conversion library: send a PDF, Word, PowerPoint, HTML, image or other file (or a URL) and get structured output back, including Markdown, HTML and the lossless DoclingDocument JSON. It parses page layout, reading order and tables, and can run OCR, with the default models bundled in the container image. It is aimed at preparing documents for search and AI pipelines rather than editing PDFs.
Pick Docling Serve when you need to turn PDFs and office documents into structured text for search or AI pipelines, and want that as an HTTP service on your own server. It wraps the Docling library: you send a file or URL and get back Markdown, HTML or the DoclingDocument JSON, with layout, reading order, tables and OCR handled by models. It is MIT licensed. Our snippet follows the README's container example with `quay.io/docling-project/docling-serve-cpu:v1.38.2` (v1.38.2 published 2026-10-27), the CPU-only image, and sets `DOCLING_SERVE_ENABLE_UI` so the web playground is served at `/ui`. The README lists that image at about 4.4 GB, so expect a long first pull. The docs say the standard image includes only the default models. The README gives no RAM minimum, and upstream's Kubernetes example requests 1 GiB with a 4 GiB limit on a GPU image, so our 4096 MB floor is an estimate for loading layout and table models on CPU. Gotenberg in this catalog is an API that converts web pages, HTML, Markdown and office formats into PDF, Stirling-PDF is a web app and API of PDF editing tools, BentoPDF runs its tools in the visitor's browser, and PDFDing manages and reads PDFs. Docling goes the other way, from documents to structured data, and does not edit PDFs.
What you need
Any VPS with at least 4096 MB of RAM
A domain you control — most self-hosted setups need HTTPS in front of them
Paid link — we earn a commission if you shop through it.
Install
Run these commands on your server:
# Docling Serve - document conversion API for PDFs and office files - adapted from the README's container example (CPU-only image, web UI enabled), pinned to v1.38.2cat > docker-compose.yml <<'EOF'services: docling-serve: image: quay.io/docling-project/docling-serve-cpu:v1.38.2 container_name: docling-serve restart: unless-stopped ports: - "5001:5001" environment: DOCLING_SERVE_ENABLE_UI: "true"EOFdocker compose up -dtimeout 400 bash -c 'until curl -fsS -o /dev/null http://localhost:5001/health; do sleep 3; done'# → http://SERVER_IP:5001/ui (web playground; the API docs are at /docs; there is no login unless you set DOCLING_SERVE_API_KEY)
Docling Serve's README and configuration docs flag these before you rely on it:
non-negotiableThere is no authentication by default: the API key (DOCLING_SERVE_API_KEY, checked as the X-Api-Key header) is empty unless you set it, and the default CORS settings allow all origins. Our snippet sets neither, so keep the port private or add a key and a reverse proxy.
non-negotiableOnly the default models ship in the image. The models guide says steps that need extra models, such as picture description or formula extraction, raise runtime errors rather than downloading them, unless you clear the artifacts path or mount a model folder.
non-negotiableThe image is large: the README lists the CPU image at 4.4 GB and the default image at 8.7 GB (amd64), with the CUDA 12.8 image at 11.4 GB. The images are hosted on quay.io and ghcr.io.
non-negotiableThe web UI is off unless DOCLING_SERVE_ENABLE_UI is true, and the /ui playground is a static app that calls the public API with the same key and policies. The API reference pages at /docs are on by default and need no key.
non-negotiableConversion results are single-use by default (DOCLING_SERVE_SINGLE_USE_RESULTS is true, removed after 300 seconds), and synchronous requests wait at most 120 seconds by default (DOCLING_SERVE_MAX_SYNC_WAIT).
non-negotiableRemote services are disabled by default (DOCLING_SERVE_ENABLE_REMOTE_SERVICES is false), so pipeline steps that call an external vision-language model API need that switched on. The quay manifest lists amd64 and arm64 for the CPU tag.
Docling Serve used 780 MB of RAM at idle on GCP e2-standard-2 · Ubuntu 26.04 LTS · Docker 29.9.3 (first boot, measured Nov 2026). We list a minimum of 4096 MB.
Is the Docling Serve install snippet verified?
The Docling Serve install snippet is marked verified on Ubuntu 26.04 (Nov 2026).