KAOS (Kelvin Agentic Operating System) is an open-source, MCP-native Python platform for building legal AI applications. The platform is structured as a modular suite of packages spanning document extraction, natural language processing, LLM orchestration, agentic workflows, and legal data sourcing, designed to run entirely within the user's own infrastructure without routing documents or prompts through a vendor cloud.

Key Features & Functions:

Document Extraction and Processing
Dedicated packages handle extraction from PDF, Word, PowerPoint, Excel, and web sources, producing a typed document abstract syntax tree (AST) in which every paragraph, table, and OCR line carries provenance metadata, including page number and bounding box. Auto re-OCR of degraded scans is supported via local ONNX or Tesseract engines.

Search and Retrieval
NLP primitives include BM25 ranking, Punkt segmentation, string distance measures, MinHash/LSH, dense embeddings, and reranking, implemented with a Rust backend. Retrieval operations return results with page and paragraph offsets preserved from the source document.

LLM Integration
A unified LLM client supports multiple providers with async operation, streaming, structured output, and per-call cost estimation. A higher-level LLM programming layer provides typed, composable, and optimisable LLM programs. The platform is LLM-agnostic and supports both local and remote models through a common interface.

Agentic Workflows
An agent runtime supports session memory, cost accounting, tiered OCR escalation, and grounded findings. Agents are configurable via recipes and patterns, with cost caps settable per matter before work begins and partial output preserved if a cap is reached mid-run.

Legal Data Sourcing
A data sourcing package provides programmatic access to government, financial, and forensic data sources including Federal Register, eCFR, EDGAR, GovInfo, GLEIF, and PACER.

Citation and Provenance
A citations package extracts, resolves, and verifies legal, financial, and accounting citations across a broad range of citation types using deterministic, AST-grounded methods. All platform outputs carry pointers back to the source page and paragraph; when the system lacks supporting evidence to answer a query, it declines rather than inferring.

Knowledge Graph Support
A graph engine provides a broad range of graph algorithms with first-class RDF and SPARQL support, enabling construction and querying of legal knowledge graphs.

Supply Chain and Security Transparency
All packages are tracked on a public compliance dashboard covering signed build attestations, software bill of materials, OpenSSF Scorecard, SLSA build provenance, and NIST SSDF mapping. The dependency tree is audited for licence compliance, with strong-copyleft licences excluded from the critical path.

Developer Tooling
A UI scaffolding package provides project templates and a React component library for building agent chat interfaces. A CLI is included for managing the platform runtime.