Xberg

Xberg is an open-source, polyglot document intelligence framework built on a native-Rust core, published under the MIT License by Kreuzberg, Inc. It turns any document, URL, code file, or audio clip into clean, structured data — text, tables, entities, and embeddings — ready to use in downstream pipelines. The project is the next iteration of Kreuzberg (previously published as kreuzberg), rebuilt and rebranded under a fresh v1 line.

What It Is

Xberg is a content-extraction engine that handles format detection, reading, OCR, transcription, and structured output in a single library call. Rather than stitching together a dozen specialized libraries, developers point Xberg at a PDF, spreadsheet, scanned image, audio file, URL, archive, or source tree and receive clean Markdown, plain text, HTML, JSON, or Djot back. The engine supports 100 document formats across 120 file extensions, code intelligence for 371 programming languages, and 6 output formats. It runs as a library, CLI tool, REST API server, MCP server, Docker container, or Helm chart — with no GPU required.

Architecture and Language Bindings

The Rust core is exposed through 15 language bindings generated by the project's own polyglot binding generator, alef. Supported runtimes include:

• Native NAPI-RS bindings for Node.js/TypeScript
• WebAssembly (@xberg-io/xberg-wasm) for browsers, Cloudflare Workers, Deno, and Bun — running at 60–80% of native speed
• Python, Go, Java/Kotlin JVM, Kotlin (Android), C#, Ruby, PHP, Elixir, Dart/Flutter, Swift, Zig, and C FFI

The CLI ships 12 commands (extract, batch, detect, formats, serve, mcp, embed, chunk, cache, api, version, completions) and can be installed via Homebrew. The Docker image at ghcr.io/xberg-io/xberg bundles the common feature set and supports API, CLI, and MCP modes.

Capabilities in Depth

Xberg's extraction pipeline covers a wide surface area:

• OCR: Tesseract, PaddleOCR, Candle (pure Rust), and VLM backends (GPT-4 Vision, Claude Vision, Gemini Vision, and 165 providers via the companion liter-llm library), with fallback chains and confidence scores.
• Audio/video transcription: Speech-to-text from MP3, M4A, WAV, WebM, and MP4 via Whisper ONNX (tiny through large-v3).
• Layout and tables: ML layout models (PP-DocLayout-V3, RT-DETR) and table structure models (TATR, SLANet) reconstruct reading order and cell grids.
• Code intelligence: Functions, classes, imports, symbols, and docstrings from 371 languages via tree-sitter, with syntax-aware chunking for RAG pipelines.
• Embeddings: Local ONNX presets (fast, balanced, quality, multilingual) or provider-hosted via 165 providers, plus cross-encoder reranking.
• Structured LLM extraction: Schema-driven JSON from any document using local engines (Ollama, LM Studio, vLLM) or remote providers — no prompt engineering required.
• Enrichment: NER (GLiNER or LLM-based), PII redaction, summarization, translation, keyword extraction (YAKE/RAKE), QR code detection, language detection, and token reduction via the TOON wire format.
• Web crawling: Auto, Document, and Crawl modes via the companion crawlberg engine.
• Archives: Recursive extraction from .zip, .tar, .gz, and .7z with zip-bomb and nesting-depth guards.

MCP Server and AI Coding Assistant Integration

Xberg ships a built-in MCP server (xberg mcp --transport stdio) exposing 9 tools, 3 prompts, and 4 resources. It integrates directly with Claude Desktop, Cursor, and other MCP-compatible agents. The project also publishes plugins for Claude Code, Codex CLI, Cursor, Gemini CLI, Factory Droid, GitHub Copilot CLI, and opencode, making document extraction available inside AI coding workflows without additional configuration.

Update: v1.0.14

The latest release is v1.0.14, published on 2026-08-04, with the repository last pushed on 2026-08-14. The project launched in January 2025 and has accumulated over 9,000 GitHub stars and 554 forks across its active development period. The README lists 16 downstream open-source projects that declare Xberg (or its predecessor kreuzberg) as a dependency, including Haystack core integrations (deepset-ai), MANSPIDER (blacklanternsecurity), and several MCP server projects. The project is part of the broader Xberg.io ecosystem, which also includes crawlberg, html-to-markdown, liter-llm, tree-sitter-language-pack, and alef.