Kreuzberg
Superseded by Xberg (2026, same author). Xberg relicenses from ELv2 to MIT/Apache 2.0, adds 4 more formats, a fifth OCR backend, three embedding types (dense/sparse/late-interaction), NER/summarization/translation/classification, double-chunking pipeline with divergence detection, and grows from 1 MCP tool to 9. The architecture and philosophy carry forward — the license is the headline change for anyone building products on it.
A polyglot document intelligence framework with a Rust core. Extracts text, metadata, and structured information from PDFs, Office documents, images, and 97+ formats. Available for basically every language you can think of, plus CLI, REST API, and MCP server.
Key Themes
Kreuzberg is the Swiss Army knife of document extraction. The breadth is staggering: 97+ file formats, 14+ language bindings, 306 programming languages for code intelligence via tree-sitter, multiple OCR backends including VLM-based OCR through GPT-4o/Claude/Gemini.
The MCP server deployment option is what makes this interesting for agent workflows. An AI agent can use Kreuzberg as a tool to ingest any document format -- PDFs, emails, archives, academic papers -- and get structured text back. This is the plumbing that makes Agentic Coding practical for document-heavy workflows.
The TOON wire format claiming 30-50% fewer tokens than JSON for RAG pipelines is a bold optimization. If your pipeline processes thousands of documents, that token savings compounds fast.
Critical Analysis
The Elastic License 2.0 is a yellow flag. ELv2 means you can't offer Kreuzberg as a managed service -- fine for internal use, but it limits how you can build products on top of it. Its successor Xberg addresses this directly by switching to MIT/Apache 2.0 dual licensing, removing the managed-service restriction. Compare with Dolphin (document parsing) which is more narrowly focused but potentially more permissively licensed for production use.
The breadth-vs-depth tradeoff is real. Supporting 97+ formats means some will be better than others. For critical production use, you'd want to benchmark Kreuzberg against specialized tools for your specific format (e.g., Dolphin for scanned documents, dedicated PDF extractors for digital PDFs).
For digital-born PDFs, pdf-inspector is exactly that dedicated extractor: one format, one dependency (lopdf), no OCR, tuned to win on tables and reading order where generic engines lose.
Still, for a "just extract text from whatever the user throws at me" use case, nothing else comes close to this coverage.
Sources: Kreuzberg Last updated: 2026-05-14