Mdkit vs Docling

Docling is an open-source Python library and toolkit for document conversion — strong PDF layout analysis, table structure recognition and OCR, with Markdown and a rich JSON document model as outputs. You install and run it yourself (it downloads and runs ML models locally, so it wants real CPU/GPU and memory).

This is not an either/or from Mdkit's side: Mdkit's async lane runs Docling as its high-fidelity engine. Mdkit is the hosted, operated version of the same capability — an API key instead of a deployment.

When Docling fits

You want the library: full control over versions, models and tuning, data that must never leave your infrastructure, no per-call cost, and you're set up to operate Python ML workloads yourself. As open source, it's free.

When Mdkit fits

You want the RESULT of Docling without operating it: one POST /v1/convert/jobs call, polling or a signed webhook, Markdown and JSON outputs from storage — plus a fast synchronous lane for small files and an MCP server for agents. Free tier to evaluate, subscription pricing for volume. Sign in with a magic link to get the API key the async lane needs — no password to choose or store.

FAQ

Is Mdkit just a hosted Docling?

For the async high-fidelity lane, essentially yes — Mdkit runs Docling as its engine there. What Mdkit adds is the operational shell: an API, keys, metered credits, a job queue with webhooks, and an MCP server.

Should I self-host Docling instead?

If you can run it, yes — it is MIT-licensed and free, and you avoid a per-conversion cost. Docling's models are large and conversion is CPU-hungry, so the real question is whether you want to operate that.

Is output identical to running Docling myself?

It is the same engine, so the conversion is the same. Mdkit pins a version and operates it for you.