Compare

Where KAOS fits.

KAOS gets compared against two different groups, and the comparison is different in each one. The proprietary legal-AI platforms (Harvey, CoCounsel, Lexis+ Protégé, Westlaw Advantage, Everlaw, Luminance, Ironclad Jurist) sell finished workflows hosted in their cloud. The open-source agentic frameworks (DSPy, LangChain, LangGraph, Pydantic-AI, Mirascope, Instructor) sell developer toolkits with no legal grounding. KAOS sits between them.

Three axes, three groups

Where each group lands.

The detailed tables below score each named vendor on a long list of capabilities. Stepping back: only one of the three groups scores on all three of open source, MCP-native by construction, and legal-domain grounding at once.

Open source Apache-2.0 MCP-native protocol = type system Legal grounding recipes + fixtures Open source closed cloud MCP-native no native surface Legal grounding finished workflows Open source permissive licenses MCP-native tool consumer only Legal grounding general-purpose
KAOS
Proprietary platforms
OSS frameworks

Reading row by row: KAOS scores on all three axes by construction. The proprietary platforms ship deep legal workflows but stay closed and have no native MCP surface. The OSS frameworks are open and Pythonic but were not built around legal documents, citations, or recipes — and they sit on the consumer side of MCP, not the server side.

Against proprietary platforms

Open source where they are closed.

The incumbents sell finished workflows hosted on their cloud, with the data inside their walls. KAOS ships the building blocks (open-source, self-hostable, served over the Model Context Protocol) that a firm or a vendor can assemble into the same workflows on its own terms.

Capability KAOSHarveyCoCounselLexis+WestlawEverlawLuminanceIronclad
Open source
Apache-2.0NoNoNoNoNoNoNo
MCP-native
Yes — to_mcp_dict() built inNoNoNoNoNoNoNo
Provenance to source span
Yes — page + bbox + char_spanYesYesYesYesYesYesYes
Refusal when uncertain
GroundedAnswer[T] schemaConfidence-backedConfidence-backedConfidence-backedDocument Analyzer flagsYes — gold standardYesYes
Citation verification
VLAIR Capability 3
kaos-citations · 8 kinds, 8 resolvers, NLICitation checkerYesShepard's Citation AgentLitigation Document AnalyzerYesYesYes
EDGAR research workflows
VLAIR #7
kaos-source — direct APIDeep ResearchYes (federated)Workflow builderSEC integrationPartialCompliance scansNo
Bring your own corpus
Yes — user owns dataVaultFederatedWorkflowsLimitedeDiscovery onlyYesCLM only
Self-host on prem
Yes — pip installNoNoNoNoNoNoNo
Customer can extend with new tools
Yes — write a KaosToolWorkflow Builder (no code)Custom workflowsCustom workflow builderLimitedNoNoNo

Against open-source frameworks

Legal grounding where they are general.

The OSS frameworks Python developers reach for first. KAOS is closer in spirit to DSPy than to any other; the head-to-head is on the LLM surface page.

When to pick which. Pick DSPy when the deepest optimizer set (GEPA, SIMBA, multi-predictor MIPROv2) and the largest community matter most. Pick KAOS when you need crash-safe batch runs with cost caps, typed programs served as MCP tools, or answers grounded in a document AST with citation verification. Use both together by calling KAOS programs from a DSPy module over MCP — KAOS is itself an MCP server, so any MCP-aware client can call it.
Capability KAOSDSPyLangChainLangGraphPydantic-AIMirascopeInstructor
Typed Signatures / Programs
Yes — DSPy-descendantYes — originalPartialPartialYesYesOutput-only
Optimizer surface
10 (Bootstrap, MIPROv2, …)14+ (deepest)NoNoNoNoNo
MCP server out of the box
Yes — kaos-mcp + 14 serversTool consumer onlyTool consumer onlyTool consumer onlyNoNoNo
AST-grounded outputs
Cited[T] · Answer[T] · GroundedAnswer[T]NoDocument loadersDocument loadersNoNoNo
Crash-safe batch with workspace
Yes — Program v3 envelope + SQLiteEvaluate (eval-shaped)ManualManualManualManualManual
Recipes for legal work
Yes — 11 named legal recipesNoGeneric templatesGeneric templatesNoNoNo
Citation extraction + verification
kaos-citations · 8 kinds + NLINoNoNoNoNoNo
Document AST with provenance
kaos-content · 34 node typesNoString chunksString chunksNoNoNo
Rust+PyO3 BM25 / tokenizer / NLP routines
kaos-nlp-core · <600 µs / 2-termNoPython BM25Python BM25NoNoNo

The lesson the market just taught

A thin layer over a model API is not a durable business.

In late 2025, Robin AI, a well-funded contract-review startup, sold its managed services to Scissero; the engineering team was acquired by Microsoft in January 2026. The lesson the market drew: durable value lives in workflow depth, in answers traceable back to evidence, and in the work of integrating with real systems. Access to the model is not enough. Thin contract-review wrappers could not survive once the underlying model APIs caught up.

KAOS is the open-source answer. Every answer traces back to a page and a bounding box. Every dependency has been license-audited, with no AGPL exposure. The whole platform speaks the Model Context Protocol, so Claude Code, Codex CLI, Gemini CLI, VS Code, and Cursor can call it directly. The recipes are named for real legal work: merger agreements, leases, court opinions, EDGAR research. Not chatbot demos.

See /why-kaos for the architectural argument and /runtime, /extraction, /llm, /agentic, /legal-intelligence, and /search-and-retrieval for what each layer does.