OPEN SOURCE
Built in the Open
Kompile is open source under the Apache 2.0 license. The entire stack — CLI, 40+ application modules, model server, 25-pass graph optimizer, MCP tool server, 20+ data connectors, and the Eclipse DL4J ML runtime — is available for you to inspect, build, and contribute to.
The Repository
One monorepo. Everything from the CLI to the native ML runtime lives in a single repository with a unified build.
100+
Maven Modules
Apache 2.0
License
44+
MCP Tools
4
Languages (Java, Rust, C++, Python)
Stack Architecture
Kompile is organized in layers. Each layer is a collection of modules you can build and run independently.
CLI & MCP Orchestration
GraalVM native binary with 100+ commands. Project init, crawl, chat, model run, agent workflows, and an MCP stdio server exposing 44+ tools to Claude Code, Codex, Gemini CLI, and Qwen.
kompile-clikompile-cli-mainkompile-cli-commonkompile-cli-plugin-apiMCP stdio & SSE transportUnix socket daemonApplication Layer (40+ Modules)
Spring Boot framework with swappable providers for LLMs (OpenAI, Anthropic, Gemini), embeddings, vector stores, guardrails, agents, and Agent-to-Agent (A2A) protocol.
kompile-app-corekompile-llms (4 providers)kompile-embeddings (5 backends)kompile-vectorstores (4 backends)kompile-guardrailskompile-a2akompile-react-agentData Sources & Loaders
20+ connectors for ingesting enterprise data. PDF, Office, email, web crawling, and cloud platform integrations with OAuth2 authentication.
kompile-loader-pdf-extendedkompile-loader-officekompile-loader-webkompile-source-confluencekompile-source-jirakompile-source-slackkompile-source-notionkompile-source-google-workspaceKnowledge Graphs & Reasoning
Full GraphRAG pipeline with multi-entity Bayesian networks (MEBN), causal inference, 9 graph maintenance primitives, 30+ MCP graph tools, TransE/RotatE embeddings, and 9 export formats.
kompile-knowledge-graphkompile-graph-algorithmskompile-graph-neo4jkompile-graph-change-trackingkompile-event-attribution (MEBN)kompile-tool-graph (MCP)kompile-evaluationModel Server & Staging
Full model lifecycle: download from HuggingFace, convert, optimize via graph compiler, train with PEFT/alignment, and serve with an OpenAI-compatible API. Multi-GPU scheduling and paged KV cache.
kompile-model-stagingkompile-model-managerkompile-sdk-servingkompile-kvcachekompile-pipelines-frameworkModel Importers & Formats
Import models from any major framework. Native GGUF/GGML parser supports LLaMA, Mistral, Phi, Qwen, Gemma, Falcon, and more with automatic dequantization.
kompile-model-importer-onnxkompile-model-importer-tensorflowkompile-model-importer-kerasGGUF/GGML native importerSafeTensors importerHuggingFace AutoModelML Runtime (Eclipse DL4J)
High-performance ML runtime with SameDiff automatic differentiation, 25-pass graph optimizer, and hardware-specific backends. Supports CUDA, cuDNN, Triton GPU compilation, llama.cpp, and ARM/Apple acceleration.
SameDiff (auto-diff engine)ND4J (tensor operations)LibND4J (native C++ backend)CUDA / cuDNN / Tritonllama.cpp helperCPU (AVX2, AVX512, NEON)OCR & Vision Models
Document understanding pipeline with detection, recognition, table extraction, and layout analysis. SameDiff VLM models for on-device OCR.
kompile-ocr-corekompile-ocr-modelsSmolDocling-256MFlorence-2Nougat / DonutCompute Graphs & Workflows
Pluggable workflow engines for orchestrating multi-step pipelines. Integrates with external workflow tools or runs standalone.
kompile-compute-graph-camelkompile-compute-graph-n8nkompile-compute-graph-droolskompile-compute-graph-xircuitskompile-compute-graph-excelSDX Mobile & Edge Runtime
Deploy optimized .sdz model bundles to Android (AAR), iOS (XCFramework), and desktop. Platform-specific acceleration with NNAPI, ARM Compute, and Metal.
Android (arm, arm64, x86)iOS (arm64, simulator)Linux / macOS / WindowsNNAPI / Metal / ARM ComputeTokenizers & NLP
Rust-based HuggingFace tokenizer bindings via JavaCPP JNI for high-performance tokenization across the pipeline.
tokenizers-rust (JNI)OpenNLP integrationLanguage detectionSentence splitting (30+ langs)Technology Stack
The full set of technologies, frameworks, and libraries that make up Kompile.
Languages
- Java 17
- Rust (tokenizers)
- C++ (LibND4J)
- Python (bridges)
Build & Deploy
- Apache Maven
- GraalVM Native Image
- Docker / OCI
- SDX mobile bundles
Frameworks
- Spring Boot 3
- Spring AI
- Picocli (CLI)
- Apache Camel
- Vert.x (serving)
ML / DL
- Eclipse DL4J
- SameDiff
- ND4J
- llama.cpp
- Triton GPU compiler
Search
- Anserini
- Apache Lucene
- BM25 + HNSW
- pgvector
- Vespa
- ChromaDB
Graph
- Neo4j
- Native graph engine
- TransE / RotatE
- Community detection
Database
- PostgreSQL
- pgvector
- Hibernate ORM
Model Formats
- GGUF/GGML
- ONNX
- TensorFlow
- Keras
- SafeTensors
- SameDiff (.sdz)
GPU & Accelerators
- CUDA / cuDNN
- Triton
- llama.cpp
- ZLUDA (AMD)
- OneDNN
- Metal / MPS
- NNAPI
Agents & MCP
- MCP stdio & SSE
- Agent-to-Agent (A2A)
- KClaw agent runner
- Enforcer guardrails
Data Sources
- Confluence
- Jira
- Slack
- Notion
- Google Workspace
- OneDrive
Optimizations
- 25-pass graph optimizer
- Attention fusion
- RMSNorm fusion
- FP16/INT8 quantization
- Speculative decoding
Repositories
Kompile is built on top of a proven deep learning foundation. Both projects are open source and actively maintained.
GetKompile/kompile
Self-hosted AI platform for RAG, agentic chat, knowledge graphs, and model inference. GraalVM native binary with 100+ modules, 44+ MCP tools, 20+ data connectors, and OpenAI-compatible serving — all in one monorepo.
View repository →
deeplearning4j/deeplearning4j
The deep learning foundation. Eclipse Deeplearning4j, ND4J, SameDiff, DataVec, and the native LibND4J C++ runtime that powers Kompile's ML engine.
View repository →
Contribute
Both projects are built by a community of contributors. Whether you're adding a new data source connector, improving GPU scheduling, or fixing a bug in the pipeline — all contributions are welcome.