Skip to content
Learn KAOS

What is KAOS?

KAOS — the Kelvin Agentic OS — is open agentic infrastructure for legal and financial work, built by 273 Ventures and published as a family of small, composable kaos-* packages under github.com/273v.

It is not a single framework you adopt wholesale. It is a set of building blocks that share one type system and one runtime, so they compose cleanly:

  • A runtime and tool model (kaos-core) — an MCP-native type system, a dependency-injection runtime, a virtual filesystem, and an artifact store that every other package builds on.
  • A document model (kaos-content) — one Block/Inline AST with provenance, so a PDF, a DOCX, a web page, and a spreadsheet all become the same shape.
  • Ingestion (kaos-pdf, kaos-office, kaos-tabular, kaos-source, kaos-web) — turn real documents and data sources into that AST.
  • LLM programming (kaos-llm-client, kaos-llm-core) — provider-native clients plus typed, composable, optimizable LLM “programs”.
  • Agents (kaos-agents) — a stateful agent runtime with memory, patterns, permissions, cost accounting, and grounded-citation findings.
  • A deterministic substrate (kaos-nlp-core, kaos-graph, kaos-citations, kaos-ml-core, kaos-nlp-transformers, kaos-names) — fast, offline NLP, graph, citation, and ML primitives.
  • Apps and serving (kaos-ui, kaos-mcp) — scaffold user-facing apps and expose any runtime over the Model Context Protocol.

Two audiences

Section titled “Two audiences”
  • Python developers build pipelines and apps directly against the packages.
  • AI agents consume the same capabilities over MCP — every package can serve its tools to Claude Code, Codex, or any MCP client.

How to learn it here

Section titled “How to learn it here”

This site teaches KAOS by running it. The recommended order — the golden path — starts simple and deterministic, then layers on LLMs and agents:

  1. Run your first example (no key, ~10 seconds)
  2. Install the toolchain
  3. Author and run a tool, build a document, call a model, and build an agent — the guided tutorials (landing milestone by milestone).

Everything you run on the spine works offline: deterministic packages run as themselves, and LLM steps use a built-in fake model so you never need an API key to learn.